# ModelFit: Full Model Catalog

> The complete ModelFit catalog in one fetch: 156 models across 27 families. 116 run locally via Ollama on Apple Silicon, NVIDIA or AMD hardware; 40 are cloud-only APIs tracked for comparison.
> Free to cite with attribution to ModelFit (modelfit.io), CC BY 4.0. Compact index: https://modelfit.io/llms.txt. Structured JSON export: https://modelfit.io/api/dataset/. Hugging Face mirror (load_dataset ready): https://huggingface.co/datasets/modelfit/modelfit-hardware-dataset. Methodology and estimate policy: https://modelfit.io/about/.
> Ollama commands are registry-verified tags. Loaded sizes and minimum memory are ModelFit engine estimates (weights + runtime overhead), not benchmarks. Sub-3-bit quants are never recommended.

## Local models (run via Ollama)

### DeepSeek
- [DeepSeek-R1 671B](https://modelfit.io/models/deepseek-r1-671b/): 671B params, 37B active MoE, Q4_K_M, ~380GB loaded, min 512GB device memory. Best for: Reasoning, Coding. Run: `ollama run deepseek-r1:671b-q4_K_M`. Requires 400GB+ RAM. Runs on Mac Studio M3 Ultra 512GB only. Source: https://huggingface.co/deepseek-ai/DeepSeek-R1.
- [DeepSeek-R1 Distill Llama 70B](https://modelfit.io/models/deepseek-r1-distill-llama-70b/): 70B params, Q4_K_M, ~42GB loaded, min 64GB device memory. Best for: Reasoning, Quality. Run: `ollama run deepseek-r1:70b`. Large distilled DeepSeek-R1 for maximum reasoning quality. Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B.
- [DeepSeek-R1 Distill Qwen 14B](https://modelfit.io/models/deepseek-r1-distill-qwen-14b/): 14B params, Q4_K_M, ~11GB loaded, min 16GB device memory. Best for: Reasoning, Quality. Run: `ollama run deepseek-r1:14b`. Larger distilled DeepSeek-R1 with enhanced reasoning capabilities. Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B.
- [DeepSeek-R1 Distill Qwen 14B (Q8)](https://modelfit.io/models/deepseek-r1-distill-qwen-14b/): 14B params, Q8_0, ~14.6GB loaded, min 24GB device memory. Best for: Reasoning, Quality. Run: `ollama run deepseek-r1:14b-qwen-distill-q8_0`. Q8_0 variant of DeepSeek-R1 Distill Qwen 14B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/DeepSeek-R1-Distill-Qwen-14B-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B.
- [DeepSeek-R1 Distill Qwen 7B](https://modelfit.io/models/deepseek-r1-distill-qwen-7b/): 7B params, Q4_K_M, ~5.5GB loaded, min 8GB device memory. Best for: Reasoning, Coding. Run: `ollama run deepseek-r1:7b`. Distilled DeepSeek-R1 reasoning model based on Qwen. Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B.
- [DeepSeek-R1 Distill Qwen 7B (Q8)](https://modelfit.io/models/deepseek-r1-distill-qwen-7b/): 7B params, Q8_0, ~7.5GB loaded, min 16GB device memory. Best for: Reasoning, Coding. Run: `ollama run deepseek-r1:7b-qwen-distill-q8_0`. Q8_0 variant of DeepSeek-R1 Distill Qwen 7B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B.

### GPT-OSS
- [GPT-OSS 120B](https://modelfit.io/models/gpt-oss-120b/): 117B params, 5.1B active MoE, MXFP4, ~65.4GB loaded, min 96GB device memory. Best for: Reasoning, Coding, Agents. Run: `ollama run gpt-oss:120b`. OpenAI open-weight MoE (Aug 2025), Apache 2.0: 117B total, 5.1B active per token (per official HF card). Native MXFP4, 65GB via Ollama, strong reasoning and tool use on 96GB+ machines. Source: https://huggingface.co/openai/gpt-oss-120b.
- [GPT-OSS 20B](https://modelfit.io/models/gpt-oss-20b/): 21B params, 3.6B active MoE, MXFP4, ~13.8GB loaded, min 24GB device memory. Best for: Chat, Coding, Reasoning. Run: `ollama run gpt-oss:20b`. OpenAI open-weight MoE (Aug 2025), Apache 2.0: 21B total, 3.6B active per token (per official HF card). Native MXFP4, 14GB via Ollama, fits 24GB machines with headroom. Source: https://huggingface.co/openai/gpt-oss-20b.

### Gemma
- [Gemma 4 31B](https://modelfit.io/models/gemma4-31b/): 31B params, Q4_K_M, ~20GB loaded, min 32GB device memory. Best for: Quality, Coding, Multimodal. Run: `ollama run gemma4:31b`. Apr 2, 2026 release from Google DeepMind. Dense 31B. MMLU-Pro 85.2, GPQA Diamond 84.3. Apache 2.0. Multimodal text+image. Source: https://huggingface.co/google/gemma-4-31B.
- [Gemma 4 31B (Q8)](https://modelfit.io/models/gemma4-31b/): 31B params, Q8_0, ~30.9GB loaded, min 48GB device memory. Best for: Quality, Coding, Multimodal. Run: `ollama run gemma4:31b-it-q8_0`. Q8_0 variant of Gemma 4 31B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/google_gemma-4-31b-it-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/google/gemma-4-31B.
- [Gemma 2 27B Instruct](https://modelfit.io/models/gemma2-27b/): 27B params, Q4_K_M, ~21GB loaded, min 32GB device memory. Best for: Quality, Coding. Run: `ollama run gemma2:27b-instruct-q4_K_M`. Excellent output quality with high memory cost. Source: https://huggingface.co/google/gemma-2-27b-it.
- [Gemma 3 27B Instruct](https://modelfit.io/models/gemma3-27b/): 27B params, Q4_K_M, ~21GB loaded, min 32GB device memory. Best for: Quality, Coding. Run: `ollama run gemma3:27b`. Large Gemma3 variant with excellent output quality. Source: https://huggingface.co/google/gemma-3-27b-it.
- [Gemma 4 26B-A4B](https://modelfit.io/models/gemma4-26b-a4b/): 26B params, 4B active MoE, Q4_K_M, ~16GB loaded, min 24GB device memory. Best for: Chat, Coding, Multimodal. Run: `ollama run gemma4:26b`. Apr 2026 MoE, named 26B-A4B but 25.2B total / 3.8B active per Google's spec table. MMLU-Pro 82.6, GPQA Diamond 82.3. Apache 2.0. Strong throughput on 24GB Mac via MLX. Source: https://huggingface.co/google/gemma-4-26B-A4B.
- [Gemma 4 26B-A4B (Q8)](https://modelfit.io/models/gemma4-26b-a4b/): 26B params, 4B active MoE, Q8_0, ~28.1GB loaded, min 48GB device memory. Best for: Chat, Coding, Multimodal. Run: `ollama run gemma4:26b-a4b-it-q8_0`. Q8_0 higher precision variant of Gemma 4 26B-A4B, near lossless quality versus the Q4 default. Prefer it with 48GB or more RAM headroom for the 28.1GB weight file. Tag registry-verified 200. Source: https://huggingface.co/google/gemma-4-26B-A4B.
- [Gemma 3 12B Instruct](https://modelfit.io/models/gemma3-12b/): 12B params, Q4_K_M, ~9.5GB loaded, min 16GB device memory. Best for: Chat, Quality. Run: `ollama run gemma3:12b`. Mid-range Gemma3 with balanced performance. Source: https://huggingface.co/google/gemma-3-12b-it.
- [Gemma 4 12B](https://modelfit.io/models/gemma4-12b/): 12B params, Q4_K_M, ~8GB loaded, min 12GB device memory. Best for: Chat, Coding, Multimodal. Run: `ollama run gemma4:12b`. Jun 3, 2026 Google DeepMind release. Unified encoder-free dense 12B, multimodal (text/image/audio), 256K context. MMLU-Pro 77.2, GPQA Diamond 78.8, LiveCodeBench v6 72.0. Beats Gemma 3 27B and runs comfortably on a 16GB Mac. Open weights. Source: https://huggingface.co/google/gemma-4-12b.
- [Gemma 4 12B (Q8)](https://modelfit.io/models/gemma4-12b/): 12B params, Q8_0, ~12.8GB loaded, min 24GB device memory. Best for: Chat, Coding, Multimodal. Run: `ollama run gemma4:12b-it-q8_0`. Q8_0 higher precision variant of Gemma 4 12B, near lossless quality versus the Q4 default. Prefer it with 24GB or more RAM headroom for the 12.8GB weight file. Tag registry-verified 200. Source: https://huggingface.co/google/gemma-4-12b.
- [Gemma 2 9B Instruct](https://modelfit.io/models/gemma2-9b/): 9B params, Q4_K_M, ~7GB loaded, min 12GB device memory. Best for: Chat, Coding. Run: `ollama run gemma2:9b-instruct-q4_K_M`. High quality on mid-to-high RAM machines. Source: https://huggingface.co/google/gemma-2-9b-it.
- [Gemma 4 E4B](https://modelfit.io/models/gemma4-e4b/): 4.5B params, Q4_K_M, ~4GB loaded, min 6GB device memory. Best for: On-device, Mobile, Chat. Run: `ollama run gemma4:e4b`. Apr 2026 on-device Gemma 4 with Per-Layer Embeddings. 4.5B effective params. Runs on 8GB Macs and high-end iPhones. Apache 2.0. Source: https://huggingface.co/google/gemma-4-E4B.
- [Gemma 4 E4B (Q8)](https://modelfit.io/models/gemma4-e4b/): 4.5B params, Q8_0, ~7.5GB loaded, min 16GB device memory. Best for: On-device, Mobile, Chat. Run: `ollama run gemma4:e4b-it-q8_0`. Q8_0 variant of Gemma 4 E4B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/google_gemma-4-e4b-it-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/google/gemma-4-E4B.
- [Gemma 3 4B Instruct](https://modelfit.io/models/gemma3-4b/): 4B params, Q4_K_M, ~3.5GB loaded, min 6GB device memory. Best for: Chat, Coding. Run: `ollama run gemma3:4b`. Latest Gemma3 with improved reasoning and efficiency. Source: https://huggingface.co/google/gemma-3-4b-it.
- [Gemma 3 4B Instruct (Q8)](https://modelfit.io/models/gemma3-4b/): 4B params, Q8_0, ~3.9GB loaded, min 8GB device memory. Best for: Chat, Coding. Run: `ollama run gemma3:4b-it-q8_0`. Q8_0 variant of Gemma 3 4B Instruct. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the unsloth/gemma-3-4b-it-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/google/gemma-3-4b-it.
- [Gemma 4 E2B](https://modelfit.io/models/gemma4-e2b/): 2.3B params, Q4_K_M, ~2.3GB loaded, min 4GB device memory. Best for: IoT, Mobile, Edge. Run: `ollama run gemma4:e2b`. Apr 2026 ultra-light Gemma 4. 2.3B effective params per Google's card. Among the fastest builds in the 2026 batch on Apple Silicon. iPhone-class device target. Source: https://huggingface.co/google/gemma-4-26B-A4B.
- [Gemma 4 E2B (Q8)](https://modelfit.io/models/gemma4-e2b/): 2.3B params, Q8_0, ~4.6GB loaded, min 8GB device memory. Best for: IoT, Mobile, Edge. Run: `ollama run gemma4:e2b-it-q8_0`. Q8_0 variant of Gemma 4 E2B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/google_gemma-4-e2b-it-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/google/gemma-4-26B-A4B.
- [Gemma 2 2B Instruct](https://modelfit.io/models/gemma2-2b/): 2B params, Q4_K_M, ~1.8GB loaded, min 4GB device memory. Best for: Chat. Run: `ollama run gemma2:2b-instruct-q4_K_M`. Small and very responsive for simple tasks. Source: https://huggingface.co/google/gemma-2-2b-it.
- [Gemma 3 1B Instruct](https://modelfit.io/models/gemma-3-1b/): 1B params, Q4_K_M, ~1GB loaded, min 2GB device memory. Best for: Chat, Mobile. Run: `ollama run gemma3:1b`. Latest Gemma generation with improved efficiency for mobile use. Source: https://huggingface.co/google/gemma-3-1b-it.
- [Gemma 3 1B Instruct (Q8)](https://modelfit.io/models/gemma-3-1b/): 1B params, Q8_0, ~1GB loaded, min 4GB device memory. Best for: Chat, Mobile. Run: `ollama run gemma3:1b-it-q8_0`. Q8_0 variant of Gemma 3 1B Instruct. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the unsloth/gemma-3-1b-it-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/google/gemma-3-1b-it.

### Granite
- [Granite 4.2 30B](https://modelfit.io/models/granite4.2-30b): 29.3B params, Q4_K_M, ~16.5GB loaded, min 24GB device memory. Best for: Coding, Chat. Run: `ollama run granite4.2:30b`. IBM Granite 4.2 flagship dense model (Aug 2026), Apache 2.0. 29B with 131K context. The :30b pull is listed at 18GB on the Ollama library page, so it wants a 24GB machine. Tag registry-verified. Source: https://huggingface.co/ibm-granite/granite-4.2-30b.
- [Granite 4.2 8B](https://modelfit.io/models/granite4.2-8b): 8.8B params, Q4_K_M, ~5GB loaded, min 8GB device memory. Best for: Chat, Coding. Run: `ollama run granite4.2:8b`. IBM Granite 4.2 mid-size instruct model (Aug 2026), Apache 2.0. Dense 8.8B with 131K context. Ollama pull about 5 GB, comfortable on 8GB machines. Tag registry-verified. Source: https://huggingface.co/ibm-granite/granite-4.2-8b.
- [Granite 4.1 8B Instruct](https://modelfit.io/models/granite4.1-8b): 8B params, Q4_K_M, ~5.5GB loaded, min 8GB device memory. Best for: Enterprise assistant, tool calling, instruction following. Run: `ollama run granite4.1:8b`. IBM dense 8B (June 2026), Apache 2.0. MMLU 73.84 per IBM model card. 131K context, OpenAI-style tool calling, 12 languages. Safe general-purpose pick for a 16GB Mac. Source: https://huggingface.co/ibm-granite/granite-4.1-8b.
- [Granite 4.2 3B](https://modelfit.io/models/granite4.2-3b): 3.7B params, Q4_K_M, ~2.1GB loaded, min 4GB device memory. Best for: Chat, Coding. Run: `ollama run granite4.2:3b`. IBM Granite 4.2 small instruct model (Aug 2026), Apache 2.0. Dense 3.7B with 131K context. Ollama pull is about 2 GB; runs on anything. Tag registry-verified. Source: https://huggingface.co/ibm-granite/granite-4.2-3b.
- [Granite 4.1 3B Instruct](https://modelfit.io/models/granite4.1-3b): 3B params, Q4_K_M, ~2GB loaded, min 4GB device memory. Best for: Lightweight chat, classification, edge tasks. Run: `ollama run granite4.1:3b`. IBM dense 3B (June 2026), Apache 2.0. MMLU 67.02 per IBM model card. 131K context, 12 languages. Safe pick for any Apple Silicon Mac, including 8 GB machines. Source: https://huggingface.co/ibm-granite/granite-4.1-3b.

### LFM2
- [LFM2 24B-A2B Instruct](https://modelfit.io/models/lfm2-24b-a2b/): 24B params, 2B active MoE, Q4_K_M, ~14GB loaded, min 24GB device memory. Best for: Local AI agents, privacy-first tool calling, MCP workflows. Run: `ollama run lfm2:24b-a2b`. LiquidAI flagship hybrid MoE (24B total, 2.3B active per the model card). Powers the LocalCowork open-source agent app over MCP. 112 tok/s on AMD CPU, 293 tok/s on H100, per LiquidAI. Requires 16GB+ unified memory. Best for privacy-sensitive enterprise workflows with zero cloud. Source: https://huggingface.co/LiquidAI/LFM2-24B-A2B.
- [LFM2.5 8B-A1B](https://modelfit.io/models/lfm2.5-8b-a1b): 8.3B params, 1.5B active MoE, Q4_K_M, ~5.5GB loaded, min 8GB device memory. Best for: On-device agents, tool calling, multilingual chat. Run: `ollama run lfm2.5:8b-a1b-q4_K_M`. LiquidAI hybrid conv+GQA model (8.3B total, 1.5B active per token). IFEval 91.84, MATH500 88.76 per official HF model card. 128K context, native tool calling, 10 languages. ~5 GB at Q4_K_M, runs on most Apple Silicon Macs. Not its strongest area: heavy coding. Source: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B.

### Laguna
- [Laguna S 2.1](https://modelfit.io/models/laguna-s-2.1): 118B params, 8B active MoE, Q4_K_M, ~96GB loaded, min 128GB device memory. Best for: Agentic coding, Long-horizon tasks. Run: `ollama run laguna-s-2.1:q4_K_M`. Jul 2026 flagship coding MoE from poolside (118B total, 8B active), aimed at long-horizon agentic work. Q4 load fills a 128GB machine; the largest open coding model with a native Ollama library tag. Source: https://huggingface.co/poolside/Laguna-S-2.1.
- [Poolside Laguna XS.2](https://modelfit.io/models/laguna-xs2/): 33B params, 3B active MoE, Q4_K_M, ~23GB loaded, min 36GB device memory. Best for: Agentic coding, long-horizon software engineering. Run: `ollama run laguna-xs.2:q4_K_M`. Poolside AI open-weight (Apache 2.0) coding model, June 2026. 33B MoE / 3B active, 128K context. Per Poolside's model card: SWE-bench Verified 69.9%, SWE-bench Multilingual 57.7%, SWE-bench Pro 46.3%, Terminal-Bench 2.0 35.7%. Coding specialist (not a general chat model). Source: https://huggingface.co/poolside/Laguna-XS.2.
- [Laguna XS 2.1](https://modelfit.io/models/laguna-xs-2.1): 33B params, 3B active MoE, Q4_K_M, ~20.3GB loaded, min 32GB device memory. Best for: Agentic coding, Long-horizon tasks. Run: `ollama run laguna-xs-2.1:q4_K_M`. Jul 2026 MoE from poolside (33B total, 3B active) built for agentic coding and long-horizon work on local machines. Fast decode for its size thanks to the small active set. Source: https://huggingface.co/poolside/Laguna-XS-2.1.

### Ling
- [Ling-3.0-flash](https://modelfit.io/models/ling-3.0-flash): 124B params, 5.1B active MoE, Q4_K_M, ~77.8GB loaded, min 128GB device memory. Best for: Agentic coding, Long context, Reasoning. Run: `undefined`. inclusionAI (Ant Group) hybrid-linear MoE, Aug 2026, MIT. Total 124B, activated 5.1B, with Kimi Delta Attention and MLA stacked 5:1, trained up to 256K context. A vision-language sibling, Ling-3.0-flash-VL, shares the backbone and gained llama.cpp support on Sept 24, 2026. No Ollama tag: the bartowski Q4_K_M GGUF is about 78 GB (128GB machine); IQ4_XS at about 69 GB squeezes onto 96GB. Source: https://huggingface.co/inclusionAI/Ling-3.0-flash.

### Llama
- [Llama 3.1 405B Instruct](https://modelfit.io/models/llama-3.1-405b): 405B params, Q4_K_M, ~243GB loaded, min 320GB device memory. Best for: Quality, Reasoning, Coding. Run: `ollama run llama3.1:405b-instruct-q4_K_M`. Requires 256GB+ RAM. Best for Mac Studio M3 Ultra and Mac Pro with large memory. Source: https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct.
- [Llama 4 Maverick](https://modelfit.io/models/llama4-maverick/): 400B params, 17B active MoE, Q4_K_M, ~245GB loaded, min 320GB device memory. Best for: Frontier quality, Long context. Run: `ollama run llama4:maverick`. Meta Llama 4 Maverick: 400B total, 17B active MoE. 1M context. Multimodal. Requires Mac Studio M3 Ultra 256GB+ for local. Source: https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct.
- [Llama 4 Scout](https://modelfit.io/models/llama4-scout/): 109B params, 17B active MoE, Q4_K_M, ~67GB loaded, min 96GB device memory. Best for: Long context, Quality, Multimodal. Run: `ollama run llama4:scout`. Meta Llama 4 Scout: 109B total, 17B active MoE. 10M context. Multimodal. Best run on Mac Studio 96GB+ or 128GB. Source: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct.
- [Llama 3.1 70B Instruct](https://modelfit.io/models/llama3.1-70b): 70B params, Q4_K_M, ~42GB loaded, min 64GB device memory. Best for: Quality, Coding. Run: `ollama run llama3.1:70b-instruct-q4_K_M`. Top quality but very hardware heavy. Source: https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct.
- [Llama 3.3 70B Instruct](https://modelfit.io/models/llama3.3-70b): 70B params, Q4_K_M, ~42GB loaded, min 64GB device memory. Best for: Quality, Coding. Run: `ollama run llama3.3:70b-instruct-q4_K_M`. Latest Llama 3.3 with improved reasoning and coding capabilities. Source: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct.
- [Llama 3.3 70B Instruct (Q8)](https://modelfit.io/models/llama3.3-70b): 70B params, Q8_0, ~75GB loaded, min 96GB device memory. Best for: Quality, Coding. Run: `ollama run llama3.3:70b-instruct-q8_0`. Q8_0 higher precision variant of Llama 3.3 70B, near lossless quality versus the Q4 default. Prefer it on 128GB or more workstations where the 75GB weight file still leaves headroom. Tag registry-verified 200. Source: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct.
- [Llama 3.3 70B Instruct (Q6)](https://modelfit.io/models/llama3.3-70b): 70B params, Q6_K, ~57.9GB loaded, min 96GB device memory. Best for: Quality, Coding. Run: `ollama run llama3.3:70b-instruct-q6_K`. Q6_K middle ground variant of Llama 3.3 70B, between the Q4 default and the near lossless Q8 sibling. Prefer it on 96GB or more workstations where the 57.9GB weight file balances quality against RAM cost. Tag registry-verified 200. Source: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct.
- [Llama 3.1 8B Instruct](https://modelfit.io/models/llama3.1-8b): 8B params, Q4_K_M, ~6.5GB loaded, min 12GB device memory. Best for: Chat, Coding. Run: `ollama run llama3.1:8b-instruct-q4_K_M`. Reliable default for local assistants. Source: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct.
- [Llama 3.1 8B Instruct (Q8)](https://modelfit.io/models/llama3.1-8b): 8B params, Q8_0, ~8GB loaded, min 16GB device memory. Best for: Chat, Coding. Run: `ollama run llama3.1:8b-instruct-q8_0`. Q8_0 variant of Llama 3.1 8B Instruct. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/Meta-Llama-3.1-8B-Instruct-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct.
- [Llama 3.1 8B Instruct (Q5)](https://modelfit.io/models/llama3.1-8b): 8B params, Q5_K_M, ~8GB loaded, min 12GB device memory. Best for: Chat, Coding. Run: `ollama run llama3.1:8b-instruct-q5_K_M`. Higher fidelity at modest speed cost. Source: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct.
- [Llama 3.2 3B Instruct](https://modelfit.io/models/llama3.2-3b): 3B params, Q4_K_M, ~2.5GB loaded, min 4GB device memory. Best for: Chat. Run: `ollama run llama3.2:3b-instruct-q4_K_M`. Fast assistant with broad ecosystem support. Source: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct.
- [Llama 3.2 1B Instruct](https://modelfit.io/models/llama3.2-1b): 1B params, Q4_K_M, ~1GB loaded, min 2GB device memory. Best for: Chat. Run: `ollama run llama3.2:1b-instruct-q4_K_M`. Ultra-fast baseline model for lightweight tasks. Source: https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct.

### MiMo
- [MiMo-V2.6-Distill-Qwen-9B](https://modelfit.io/models/mimo-v2.6-distill-qwen-9b): 9B params, Q4_K_M, ~5.8GB loaded, min 16GB device memory. Best for: Agentic coding, Small machines. Run: `undefined`. Xiaomi's distillation of MiMo-V2.6 into Qwen3.5-9B (Sept 2026, MIT). Per the card, SWE Verified 61.1 vs 60.0 for the base Qwen3.5-9B and SWE Pro 44.6 vs 32.0. No Ollama library tag; run the GGUF through llama.cpp or LM Studio (bartowski Q4_K_M is about 6 GB, so a 16GB Mac runs it comfortably). Source: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B.

### MiniCPM
- [MiniCPM-V 4.5 8B](https://modelfit.io/models/minicpm-v4.5-8b): 8.7B params, Q4_K_M, ~5.7GB loaded, min 12GB device memory. Best for: Vision, Multimodal. Run: `ollama run minicpm-v4.5:8b`. OpenBMB vision-language model (Aug 2025), Apache 2.0: 8.7B with image understanding, a strong local pick for screenshots and document parsing. Ollama pull about 6 GB. Tag registry-verified. Source: https://huggingface.co/openbmb/MiniCPM-V-4_5.

### Mistral
- [Devstral 2 123B](https://modelfit.io/models/devstral-2-123b/): 123B params, Q4_K_M, ~74.9GB loaded, min 96GB device memory. Best for: Agentic coding, Software engineering. Run: `ollama run devstral-2:123b`. Mistral's flagship coding agent (late 2025, modified MIT license) with a 256k context window. SWE Bench Verified 72.2% per Mistral's card, above its 24B sibling Devstral Small 2. It is dense, so every token reads all 123B weights: expect slow decode even on a Max or Ultra. The Ollama :123b pull is Q4_K_M at about 75 GB, which needs a 96GB machine. Tag registry-verified. Source: https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512.
- [Mixtral 8x7B Instruct](https://modelfit.io/models/mixtral-8x7b/): 46.7B params, 13B active MoE, Q4_K_M, ~30GB loaded, min 48GB device memory. Best for: Coding, Quality. Run: `ollama run mixtral:8x7b`. MoE quality boost with large memory demand. Source: https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1.
- [Mistral Small 3.1](https://modelfit.io/models/mistral-small-3.1): 24B params, Q4_K_M, ~15GB loaded, min 24GB device memory. Best for: Chat, Coding. Run: `ollama run mistral-small3.1:24b`. Mistral Small 3.1, latest iteration with improved efficiency. Runs locally on 24GB+ RAM. Source: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503.
- [Mistral Small 3.1 (Q8)](https://modelfit.io/models/mistral-small-3.1-24b): 24B params, Q8_0, ~23.3GB loaded, min 36GB device memory. Best for: Chat, Coding. Run: `ollama run mistral-small3.1:24b-instruct-2503-q8_0`. Q8_0 variant of Mistral Small 3.1. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the unsloth/Mistral-Small-3.1-24B-Instruct-2503-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503.
- [Devstral Small 2 24B](https://modelfit.io/models/devstral-small-2-24b/): 24B params, Q4_K_M, ~15.2GB loaded, min 24GB device memory. Best for: Agentic coding, Software engineering. Run: `ollama run devstral-small-2:24b`. Mistral's 24B dense coding agent (late 2025, Apache 2.0) with a 256k context window and image input. SWE Bench Verified 68.0% per Mistral's card, level with GLM 4.6 at 355B in the same table. A coding specialist, so it is kept out of the general best-pick. The Ollama :24b pull is Q4_K_M at about 15 GB, which fits a 24GB Mac. Tag registry-verified. Source: https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512.
- [Mistral Small 22B](https://modelfit.io/models/mistral-small-22b/): 22B params, Q4_K_M, ~17GB loaded, min 32GB device memory. Best for: Coding, Quality. Run: `ollama run mistral-small:22b`. Strong quality for advanced local use. Source: https://huggingface.co/mistralai.
- [Ministral 3 14B](https://modelfit.io/models/ministral-3-14b/): 13.5B params, Q4_K_M, ~9.1GB loaded, min 16GB device memory. Best for: Chat, Reasoning, Vision. Run: `ollama run ministral-3:14b`. Mistral edge model (2512 release, Apache 2.0): 13.5B language model plus a 0.4B vision encoder, 256k context, text and image input. WildBench 68.5 vs 65.1 for Qwen3 14B per Mistral's card. The Ollama :14b pull is Q4_K_M at about 9 GB. Tag registry-verified. Source: https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512.
- [Mistral Nemo 12B](https://modelfit.io/models/mistral-nemo-12b/): 12B params, Q4_K_M, ~9.5GB loaded, min 16GB device memory. Best for: Chat, Translation. Run: `ollama run mistral-nemo:12b`. Strong multilingual quality at moderate size. Source: https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407.
- [Mistral Nemo 12B (Q8)](https://modelfit.io/models/mistral-nemo-12b/): 12B params, Q8_0, ~12.1GB loaded, min 24GB device memory. Best for: Chat, Translation. Run: `ollama run mistral-nemo:12b-instruct-2407-q8_0`. Q8_0 variant of Mistral Nemo 12B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/Mistral-Nemo-Instruct-2407-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407.
- [Ministral 3 8B](https://modelfit.io/models/ministral-3-8b/): 8.4B params, Q4_K_M, ~6GB loaded, min 12GB device memory. Best for: Chat, Coding, Vision. Run: `ollama run ministral-3:8b`. Mistral edge model (2512 release, Apache 2.0): 8.4B language model plus a 0.4B vision encoder, 256k context, text and image input. WildBench 66.8 vs 66.3 for Qwen3-VL-8B-Instruct per Mistral's card. The Ollama :8b pull is Q4_K_M at about 6 GB. Tag registry-verified. Source: https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512.
- [Mistral 7B Instruct](https://modelfit.io/models/mistral-7b/): 7B params, Q4_K_M, ~5.5GB loaded, min 8GB device memory. Best for: Chat, Coding. Run: `ollama run mistral:7b-instruct-v0.3-q4_K_M`. Good all-around quality in a mid-size footprint. Source: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3.
- [Mistral 7B Instruct (Q5)](https://modelfit.io/models/mistral-7b/): 7B params, Q5_K_M, ~4.8GB loaded, min 8GB device memory. Best for: Chat, Coding. Run: `ollama run mistral:7b-instruct-v0.3-q5_K_M`. Q5_K_M variant of Mistral 7B Instruct. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/Mistral-7B-Instruct-v0.3-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3.
- [Mistral 7B Instruct (Q8)](https://modelfit.io/models/mistral-7b/): 7B params, Q8_0, ~7.2GB loaded, min 16GB device memory. Best for: Chat, Coding. Run: `ollama run mistral:7b-instruct-v0.3-q8_0`. Q8_0 variant of Mistral 7B Instruct. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/Mistral-7B-Instruct-v0.3-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3.
- [Ministral 3 3B](https://modelfit.io/models/ministral-3-3b/): 3.4B params, Q4_K_M, ~3GB loaded, min 6GB device memory. Best for: Chat, Vision, Edge. Run: `ollama run ministral-3:3b`. Mistral edge model (2512 release, Apache 2.0): 3.4B language model plus a 0.4B vision encoder, 256k context, text and image input. The Ollama :3b pull is Q4_K_M at about 3 GB. Tag registry-verified. Source: https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512.

### Muse
- [Muse Glimmer 30B](https://modelfit.io/models/muse-glimmer-30b/): 29.8B params, Q4_K_M, ~16.9GB loaded, min 32GB device memory. Best for: Chat. Run: `ollama run muse-glimmer:30b`. Community 30B general chat model (Aug 2026, meta-models org on Hugging Face). Limited public evaluation data; treat quality as unverified. Ollama pull about 17 GB, fits a 24GB machine. Tag registry-verified. Source: https://huggingface.co/meta-models/Muse-Glimmer-30B.

### Nemotron
- [Nemotron 3 Super 120B-A12B](https://modelfit.io/models/nemotron-3-super-120b/): 120B params, 12B active MoE, Q4_K_M, ~86.8GB loaded, min 128GB device memory. Best for: Agentic workflows, Long context, Reasoning. Run: `ollama run nemotron-3-super:120b`. NVIDIA's mid-size open model (March 11, 2026, NVIDIA Open Model License). 120B total with 12B active in a LatentMoE hybrid of Mamba-2, MoE and a few attention layers, plus Multi-Token Prediction. Context up to 1M tokens. The card reports MMLU-Pro 83.73 and SWE-Bench (OpenHands) 60.47. The Ollama :120b pull is Q4_K_M at about 87 GB, so it needs a 128GB machine. Tag registry-verified. Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.
- [NVIDIA Nemotron Cascade 2 30B-A3B](https://modelfit.io/models/nemotron-cascade-2-30b-a3b/): 30B params, 3B active MoE, Q6_K, ~24GB loaded, min 36GB device memory. Best for: Reasoning, Math, Agentic tasks. Run: `ollama run nemotron-cascade-2:30b`. NVIDIA release (Mar 19, 2026; Ollama tag added Jun 2026). 30B MoE with 3B active params, 256K context, thinking + instruct modes. Gold-medal performance at the 2025 IMO and IOI (IOI score 439.3) per NVIDIA's model card. The :30b tag ships ~24GB (Q6_K), so it wants a 32GB machine. Source: https://huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B.
- [Nemotron 3.5 Lightning 30B-A3B](https://modelfit.io/models/nemotron-3.5-lightning-30b): 30B params, 3B active MoE, Q4_K_M, ~23.7GB loaded, min 36GB device memory. Best for: Agentic, Coding, Long context. Run: `ollama run nemotron-3.5-lightning:30b`. NVIDIA hybrid Mamba-MoE (Aug 2026): 30B total with 3B active, 52 layers of which only 6 are full attention (2 KV heads), so KV-cache is a fraction of a dense 30B and context scales to 1M tokens. NVFP4 official build; the Ollama :30b pull is about 24 GB and wants a 36GB machine. Tag registry-verified. Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4.

### North
- [Cohere North Mini Code](https://modelfit.io/models/north-mini-code-1/): 30B params, 3B active MoE, Q4_K_M, ~19GB loaded, min 32GB device memory. Best for: Agentic coding, repo-scale software engineering. Run: `ollama run north-mini-code-1.0:q4_K_M`. Cohere's first developer-focused model (Jun 11, 2026), Apache 2.0. 30B MoE / 3B active, 256K context, up to 64K output, per Cohere's model card. Coding specialist (not a general chat model). Source: https://huggingface.co/CohereLabs/North-Mini-Code-1.0.

### Olmo
- [Olmo 3.1 32B Instruct](https://modelfit.io/models/olmo-3.1-32b): 32B params, Q4_K_M, ~19.5GB loaded, min 32GB device memory. Best for: General chat, Research, Fully open. Run: `ollama run olmo-3.1:32b`. Ai2's fully open 32B dense model (Apache 2.0): weights, data and training code are all public, which is the reason to pick it. On raw capability it trails Qwen 3 32B in no-think mode on MMLU (80.9 vs 85.8) and GPQA (48.6 vs 54.4) in Ai2's own table. The Ollama pull is Q4_K_M at about 19 GB. Tag registry-verified. Source: https://huggingface.co/allenai/Olmo-3.1-32B-Instruct.

### Ornith
- [Ornith 1.0 35B](https://modelfit.io/models/ornith-35b/): 35B params, Q4_K_M, ~21.2GB loaded, min 32GB device memory. Best for: Agentic coding. Run: `ollama run ornith:35b`. Jul 2026 MoE member of the Ornith agentic-coding family (MIT), sized for single-GPU and 32GB-Mac deployment. Publisher claims state-of-the-art coding-agent results among open models of its size. Source: https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B.
- [Ornith 1.0 9B](https://modelfit.io/models/ornith-9b/): 9B params, Q4_K_M, ~5.6GB loaded, min 8GB device memory. Best for: Agentic coding on small machines. Run: `ollama run ornith:9b`. Jul 2026 self-improving agentic coder from Deep Reinforce (MIT), post-trained on top of Gemma 4 and Qwen 3.5. Publisher reports coding-agent results near 35B-class models; a standout pick for 16GB Macs. Source: https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B.

### Phi
- [Phi-3 Medium 14B](https://modelfit.io/models/phi3-medium/): 14B params, Q4_K_M, ~11GB loaded, min 16GB device memory. Best for: Coding, Quality. Run: `ollama run phi3:14b-medium-4k-instruct-q4_K_M`. Very strong coding quality on capable systems. Source: https://huggingface.co/microsoft/Phi-3-medium-4k-instruct.
- [Phi-4 14B](https://modelfit.io/models/phi4-14b/): 14B params, Q4_K_M, ~11.5GB loaded, min 24GB device memory. Best for: Coding, Quality. Run: `ollama run phi4:14b-q4_K_M`. High reasoning quality with slower throughput. Source: https://huggingface.co/microsoft/phi-4.
- [Phi-4 14B (Q8)](https://modelfit.io/models/phi4-14b/): 14B params, Q8_0, ~14.5GB loaded, min 24GB device memory. Best for: Coding, Quality. Run: `ollama run phi4:14b-q8_0`. Q8_0 variant of Phi-4 14B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/Phi-4-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/microsoft/phi-4.
- [Phi-3 Mini 3.8B](https://modelfit.io/models/phi3-mini/): 3.8B params, Q4_K_M, ~3.2GB loaded, min 6GB device memory. Best for: Coding, Chat. Run: `ollama run phi3:3.8b-mini-4k-instruct-q4_K_M`. Fast and surprisingly capable coding assistant. Source: https://huggingface.co/microsoft/Phi-3-mini-4k-instruct.
- [Phi-4 Mini 3.8B](https://modelfit.io/models/phi4-mini-3.8b): 3.8B params, Q4_K_M, ~3.2GB loaded, min 6GB device memory. Best for: Coding, Chat. Run: `ollama run phi4-mini:3.8b`. Latest Phi-4 mini variant with enhanced reasoning capabilities. Source: https://huggingface.co/microsoft/Phi-4-mini-instruct.
- [Phi-4 Mini 3.8B (Q8)](https://modelfit.io/models/phi4-mini-3.8b): 3.8B params, Q8_0, ~3.8GB loaded, min 8GB device memory. Best for: Coding, Chat. Run: `ollama run phi4-mini:3.8b-q8_0`. Q8_0 variant of Phi-4 Mini 3.8B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the unsloth/Phi-4-mini-instruct-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/microsoft/Phi-4-mini-instruct.

### Qwen
- [Qwen3 235B A22B](https://modelfit.io/models/qwen3-235b-a22b/): 235B params, 22B active MoE, Q4_K_M, ~130GB loaded, min 192GB device memory. Best for: Quality, Reasoning. Run: `ollama run qwen3:235b-a22b-q4_K_M`. Massive MoE model with 22B active parameters. Requires 192GB+ RAM. Source: https://huggingface.co/Qwen/Qwen3-235B-A22B.
- [Qwen3.8-Flash-Next](https://modelfit.io/models/qwen3.8-flash-next): 125B params, 6B active MoE, Q4_K_XL, ~111.3GB loaded, min 192GB device memory. Best for: Agentic coding, Reasoning, Multimodal. Run: `undefined`. Qwen4 architecture preview (Aug 26, 2026), open weights under Qwen Community License 1.0. 125B params with 6B activated (512 experts, 10 routed + 1 shared), PLUS a 51B n-gram embedding table and 4B MTP layer, so GGUFs are larger than the active count suggests. Hybrid Gated DeltaNet + Qwen Sparse Attention. Native 262K context, extensible to 1M. Vision encoder. No Ollama tag; mainline llama.cpp runs it (qwen4exp architecture). The unsloth UD-Q4_K_XL GGUF is about 111 GB (192GB machine); UD-IQ4_XS at about 94 GB fits 128GB. Production API version: Qwen3.8-Flash on Qwen Cloud (1M context default). Source: https://huggingface.co/Qwen/Qwen3.8-Flash-Next.
- [Qwen3.5 122B-A10B Instruct](https://modelfit.io/models/qwen3.5-122b-a10b): 122B params, 10B active MoE, Q4_K_M, ~72GB loaded, min 96GB device memory. Best for: Frontier-level reasoning, Complex tasks. Run: `ollama run qwen3.5:122b-a10b`. MoE model (122B total, 10B active). Closes the gap between medium-sized and frontier models. Best for complex agent scenarios. Requires high-end Mac Studio. Source: https://huggingface.co/Qwen/Qwen3.5-122B-A10B.
- [Qwen3-Next 80B-A3B](https://modelfit.io/models/qwen3-next-80b-a3b/): 80B params, 3B active MoE, Q4_K_M, ~50.4GB loaded, min 72GB device memory. Best for: Chat, Coding, Long Context. Run: `ollama run qwen3-next:80b`. Qwen hybrid-attention MoE (Sep 2025): 80B total, 3B activated (per official HF card). 50GB via Ollama, the sweet spot for 72-96GB workstations and Macs. Source: https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct.
- [Qwen3-Next 80B-A3B (Q8)](https://modelfit.io/models/qwen3-next-80b-a3b/): 80B params, 3B active MoE, Q8_0, ~84.8GB loaded, min 128GB device memory. Best for: Chat, Coding, Long Context. Run: `ollama run qwen3-next:80b-a3b-instruct-q8_0`. Q8_0 higher precision variant of Qwen3-Next 80B-A3B, near lossless quality versus the Q4 default. Prefer it on 128GB or more workstations where the 84.8GB weight file still leaves headroom. Tag registry-verified 200. Source: https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct.
- [Qwen3.5 35B-A3B Instruct](https://modelfit.io/models/qwen3.5-35b-a3b): 35B params, 3B active MoE, Q4_K_M, ~20GB loaded, min 32GB device memory. Best for: Reasoning, Coding, Agent scenarios. Run: `ollama run qwen3.5:35b-a3b`. MoE model (35B total, 3B active). Surpasses Qwen3-235B-A22B. Runs on 24GB Mac/RAM. Better architecture beats bigger parameter counts. Source: https://huggingface.co/Qwen/Qwen3.5-35B-A3B.
- [Qwen3.6 35B-A3B](https://modelfit.io/models/qwen3.6-35b-a3b): 35B params, 3B active MoE, Q4_K_M, ~22GB loaded, min 32GB device memory. Best for: Reasoning, Coding, Agents. Run: `ollama run qwen3.6:35b-a3b`. Apr 2026 MoE (35B total, 3B active). Successor to Qwen3.5-35B-A3B with better reasoning and tool calling. 262K context. MoE routing keeps decode fast for its size on Apple Silicon. Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B.
- [Qwen3.6 35B-A3B (Q8)](https://modelfit.io/models/qwen3.6-35b-a3b): 35B params, 3B active MoE, Q8_0, ~38.7GB loaded, min 64GB device memory. Best for: Reasoning, Coding, Agents. Run: `ollama run qwen3.6:35b-a3b-q8_0`. Q8_0 higher precision variant of Qwen3.6 35B-A3B, near lossless quality versus the Q4 default. Prefer it with 64GB or more RAM headroom for the larger 38.7GB weight file. Tag registry-verified 200 on the ollama.ai manifest API. Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B.
- [Qwen3.5 35B-A3B Instruct (Q8)](https://modelfit.io/models/qwen3.5-35b-a3b): 35B params, 3B active MoE, Q8_0, ~38.7GB loaded, min 64GB device memory. Best for: Reasoning, Coding, Agent scenarios. Run: `ollama run qwen3.5:35b-a3b-q8_0`. Q8_0 higher precision variant of Qwen3.5 35B-A3B, near lossless quality versus the Q4 default. Prefer it with 64GB or more RAM headroom for the 38.7GB weight file. Tag registry-verified 200. Source: https://huggingface.co/Qwen/Qwen3.5-35B-A3B.
- [Qwen3 30B](https://modelfit.io/models/qwen3-30b/): 30B params, 3B active MoE, Q4_K_M, ~22GB loaded, min 32GB device memory. Best for: Quality, Coding. Run: `ollama run qwen3:30b-a3b-q4_K_M`. Large Qwen3 variant for high-quality outputs. Source: https://huggingface.co/Qwen/Qwen3-30B-A3B.
- [Qwen3 30B (Q8)](https://modelfit.io/models/qwen3-30b-a3b/): 30B params, 3B active MoE, Q8_0, ~30.3GB loaded, min 48GB device memory. Best for: Quality, Coding. Run: `ollama run qwen3:30b-a3b-q8_0`. Q8_0 variant of Qwen3 30B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/Qwen_Qwen3-30B-A3B-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/Qwen/Qwen3-30B-A3B.
- [Qwen3.5 27B Instruct](https://modelfit.io/models/qwen3.5-27b): 27B params, Q4_K_M, ~16GB loaded, min 24GB device memory. Best for: Chat, Coding, Complex reasoning. Run: `ollama run qwen3.5:27b`. Mid-size model narrowing the gap between medium and frontier models. Excellent for complex agent scenarios. Source: https://huggingface.co/Qwen/Qwen3.5-27B.
- [Qwen3.5 27B Instruct (Q8)](https://modelfit.io/models/qwen3.5-27b): 27B params, Q8_0, ~27.1GB loaded, min 48GB device memory. Best for: Chat, Coding, Complex reasoning. Run: `ollama run qwen3.5:27b-q8_0`. Q8_0 variant of Qwen3.5 27B Instruct. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/Qwen_Qwen3.5-27B-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/Qwen/Qwen3.5-27B.
- [Qwen3.6 27B](https://modelfit.io/models/qwen3.6-27b): 27B params, Q4_K_M, ~18GB loaded, min 32GB device memory. Best for: Coding, Quality, Long context. Run: `ollama run qwen3.6:27b`. Apr 2026 flagship dense Qwen. 262K context, extensible to 1M. SWE-Bench Verified 77.2, MMLU-Pro 86.2 per Qwen's model card. Apache 2.0. Strongest dense open-weight coder ModelFit tracks at this size. Source: https://huggingface.co/Qwen/Qwen3.6-27B.
- [Qwen3.8 27B](https://modelfit.io/models/qwen3.8-27b): 27B params, Q4_K_M, ~16.5GB loaded, min 24GB device memory. Best for: Coding, Agent, Vision, Long context. Run: `ollama run qwen3.8:27b`. Aug 2026 flagship dense Qwen with native vision-language. Apache 2.0. 262K context, extensible to 1M. MTP decoding. Terminal Bench 2.1 73.0, SWE-bench Pro 61.7 per Qwen card. Source: https://huggingface.co/Qwen/Qwen3.8-27B.
- [Qwen3.8 27B (Q8)](https://modelfit.io/models/qwen3.8-27b): 27B params, Q8_0, ~27.1GB loaded, min 48GB device memory. Best for: Coding, Agent, Vision, Long context. Run: `ollama run qwen3.8:27b-q8_0`. Q8_0 variant of Qwen3.8 27B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the unsloth/Qwen3.8-27B-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/Qwen/Qwen3.8-27B.
- [Qwen3.6 27B (Q8)](https://modelfit.io/models/qwen3.6-27b): 27B params, Q8_0, ~30GB loaded, min 48GB device memory. Best for: Coding, Quality, Long context. Run: `ollama run qwen3.6:27b-q8_0`. Q8_0 higher precision variant of Qwen3.6 27B, near lossless quality versus the Q4 default. Prefer it with 48GB or more RAM headroom for the 30GB weight file. Tag registry-verified 200. Source: https://huggingface.co/Qwen/Qwen3.6-27B.
- [Qwen2.5 14B Instruct](https://modelfit.io/models/qwen2.5-14b): 14B params, Q4_K_M, ~11GB loaded, min 16GB device memory. Best for: Coding, Chat. Run: `ollama run qwen2.5:14b-instruct-q4_K_M`. High quality if RAM allows. Source: https://huggingface.co/Qwen/Qwen2.5-14B-Instruct.
- [Qwen2.5 Coder 14B](https://modelfit.io/models/qwen2.5-coder-14b): 14B params, Q4_K_M, ~11GB loaded, min 16GB device memory. Best for: Coding. Run: `ollama run qwen2.5-coder:14b`. Great code generation for higher-memory machines. Source: https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct.
- [Qwen3 14B](https://modelfit.io/models/qwen3-14b/): 14B params, Q4_K_M, ~11GB loaded, min 16GB device memory. Best for: Coding, Quality. Run: `ollama run qwen3:14b-q4_K_M`. Mid-range Qwen3 with excellent coding performance. Source: https://huggingface.co/Qwen/Qwen3-14B.
- [Qwen3 14B (Q8)](https://modelfit.io/models/qwen3-14b/): 14B params, Q8_0, ~15.9GB loaded, min 24GB device memory. Best for: Coding, Quality. Run: `ollama run qwen3:14b-q8_0`. Q8_0 higher precision variant of Qwen3 14B, near lossless quality versus the Q4 default. Prefer it with 24GB or more RAM headroom for the 15.9GB weight file. Tag registry-verified 200. Source: https://huggingface.co/Qwen/Qwen3-14B.
- [Qwen3.5 9B Instruct](https://modelfit.io/models/qwen3.5-9b): 9B params, Q4_K_M, ~7GB loaded, min 12GB device memory. Best for: Quality, Coding, Reasoning. Run: `ollama run qwen3.5:9b`. 91.1 MMLU-Redux rivals 30B+ models. Native multimodal, 262K context. Near-frontier quality under 10GB RAM. Source: https://huggingface.co/Qwen/Qwen3.5-9B.
- [Qwen3.5 9B Instruct (Q8)](https://modelfit.io/models/qwen3.5-9b): 9B params, Q8_0, ~10.7GB loaded, min 16GB device memory. Best for: Quality, Coding, Reasoning. Run: `ollama run qwen3.5:9b-q8_0`. Q8_0 higher precision variant of Qwen3.5 9B, near lossless quality versus the Q4 default. Prefer it on 16GB or more RAM where the 10.7GB weight file still leaves headroom. Tag registry-verified 200. Source: https://huggingface.co/Qwen/Qwen3.5-9B.
- [Qwen3 8B](https://modelfit.io/models/qwen3-8b/): 8B params, Q4_K_M, ~6.5GB loaded, min 12GB device memory. Best for: Chat, Coding. Run: `ollama run qwen3:8b-q4_K_M`. Latest Qwen3 generation with enhanced reasoning. Source: https://huggingface.co/Qwen/Qwen3-8B.
- [Qwen3 8B (Q8)](https://modelfit.io/models/qwen3-8b/): 8B params, Q8_0, ~8.1GB loaded, min 16GB device memory. Best for: Chat, Coding. Run: `ollama run qwen3:8b-q8_0`. Q8_0 variant of Qwen3 8B. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/Qwen_Qwen3-8B-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/Qwen/Qwen3-8B.
- [Qwen2.5 7B Instruct](https://modelfit.io/models/qwen2.5-7b): 7B params, Q4_K_M, ~5.5GB loaded, min 8GB device memory. Best for: Chat, Coding. Run: `ollama run qwen2.5:7b-instruct-q4_K_M`. Balanced option for everyday coding and chat. Source: https://huggingface.co/Qwen/Qwen2.5-7B-Instruct.
- [Qwen2.5 Coder 7B](https://modelfit.io/models/qwen2.5-coder-7b): 7B params, Q4_K_M, ~5.5GB loaded, min 8GB device memory. Best for: Coding. Run: `ollama run qwen2.5-coder:7b`. Code-focused balance of speed and quality. Source: https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct.
- [Qwen3.5 4B Instruct](https://modelfit.io/models/qwen3.5-4b): 4B params, Q4_K_M, ~3.5GB loaded, min 6GB device memory. Best for: Coding, Agents, Multimodal. Run: `ollama run qwen3.5:4b`. 88.8 MMLU-Redux beats 20B-class models. Native multimodal, 262K context. Best quality-per-GB ratio in its class. Source: https://huggingface.co/Qwen/Qwen3.5-4B.
- [Qwen3.5 4B Instruct (Q8)](https://modelfit.io/models/qwen3.5-4b): 4B params, Q8_0, ~4.3GB loaded, min 8GB device memory. Best for: Coding, Agents, Multimodal. Run: `ollama run qwen3.5:4b-q8_0`. Q8_0 variant of Qwen3.5 4B Instruct. Heavier build, higher precision than the Q4 default; file size is the real GGUF from the bartowski/Qwen_Qwen3.5-4B-GGUF Hugging Face repo. Ollama tag registry-verified. Source: https://huggingface.co/Qwen/Qwen3.5-4B.
- [Qwen2.5 3B Instruct](https://modelfit.io/models/qwen2.5-3b): 3B params, Q4_K_M, ~2.5GB loaded, min 4GB device memory. Best for: Chat, Coding. Run: `ollama run qwen2.5:3b-instruct-q4_K_M`. Strong small model with better reasoning than 1.5B. Source: https://huggingface.co/Qwen/Qwen2.5-3B-Instruct.
- [Qwen3.5 2B Instruct](https://modelfit.io/models/qwen3.5-2b): 2B params, Q4_K_M, ~1.8GB loaded, min 4GB device memory. Best for: Chat, Edge tasks. Run: `ollama run qwen3.5:2b`. Fast small Qwen3.5 with native multimodal and 262K context. Great for edge tasks and quick chat. Source: https://huggingface.co/Qwen/Qwen3.5-2B.
- [Qwen2.5 1.5B Instruct](https://modelfit.io/models/qwen2.5-1.5b): 1.5B params, Q4_K_M, ~1.5GB loaded, min 4GB device memory. Best for: Chat, Translation. Run: `ollama run qwen2.5:1.5b-instruct-q4_K_M`. Very light general assistant for constrained devices. Source: https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct.
- [Qwen3.5 0.8B Instruct](https://modelfit.io/models/qwen3.5-0.8b): 0.8B params, Q4_K_M, ~0.8GB loaded, min 2GB device memory. Best for: Chat, Mobile. Run: `ollama run qwen3.5:0.8b`. Ultra-light Qwen3.5 with native multimodal. 262K context. Ideal for mobile and IoT edge devices. Source: https://huggingface.co/Qwen/Qwen3.5-0.8B.
- [Qwen2.5 0.5B Instruct](https://modelfit.io/models/qwen2.5-0.5b): 0.5B params, Q4_K_M, ~0.8GB loaded, min 2GB device memory. Best for: Chat, Mobile. Run: `ollama run qwen2.5:0.5b-instruct-q4_K_M`. Ultra-lightweight model used in Keiro, ideal for mobile devices. Source: https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct.

### SmolLM
- [SmolLM2 360M](https://modelfit.io/models/smollm2-360m/): 0.36B params, Q4_K_M, ~0.5GB loaded, min 1GB device memory. Best for: Chat, Embedded. Run: `ollama run smollm2:360m`. Extremely lightweight model for embedded devices and IoT. Source: https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct.

### Zhipu
- [GLM-4.7-Flash 30B-A3B](https://modelfit.io/models/glm-4.7-flash): 30B params, 3B active MoE, Q4_K_M, ~19GB loaded, min 32GB device memory. Best for: Coding, Agentic tasks, Reasoning. Run: `ollama run glm-4.7-flash`. Z.ai's local-class GLM (Jan 2026, MIT). A 30B-A3B MoE with compressed multi-latent attention, so decode speed tracks the 3B active path. SWE-bench Verified 59.2 per the model card, well ahead of Qwen3-30B-A3B-Thinking-2507 in the same table. The Ollama default pull is Q4_K_M at about 19 GB, so a 32GB Mac runs it with room for context. Tag registry-verified. Source: https://huggingface.co/zai-org/GLM-4.7-Flash.

## Cloud-only models (hosted APIs, tracked for comparison)

### Claude
- [Claude 3.5 Sonnet](https://modelfit.io/models/claude-3.5-sonnet): Cloud API, undisclosed size. Best for: Coding, Chat. Cloud/API only model family, not runnable locally via Ollama. Parameter count undisclosed (cloud API). Source: https://www.anthropic.com/claude.
- [Claude 3.7 Sonnet](https://modelfit.io/models/claude-3.7-sonnet): Cloud API, undisclosed size. Best for: Coding, Quality. Cloud/API only model family, not runnable locally via Ollama. Parameter count undisclosed (cloud API). Source: https://www.anthropic.com/claude.
- [Claude 3 Opus](https://modelfit.io/models/claude-opus-3/): Cloud API, undisclosed size. Best for: Quality, Analysis. Cloud/API only Opus family model; not runnable locally via Ollama. Parameter count undisclosed (cloud API). Source: https://www.anthropic.com/claude.
- [Claude 4 Opus](https://modelfit.io/models/claude-opus-4/): Cloud API, undisclosed size. Best for: Quality, Coding. Cloud/API only Opus family model; not runnable locally via Ollama. Parameter count undisclosed (cloud API). Source: https://www.anthropic.com/claude.
- [Claude 4 Sonnet](https://modelfit.io/models/claude-4-sonnet/): Cloud API, undisclosed size. Best for: Coding, Quality. Cloud/API only Claude 4 Sonnet with enhanced reasoning capabilities. Parameter count undisclosed (cloud API). Source: https://www.anthropic.com/claude.
- [Claude Opus 4.7](https://modelfit.io/models/claude-opus-4.7): Cloud API, undisclosed size. Best for: Coding, Quality, Reasoning. Apr 16, 2026 Anthropic flagship, superseded by Opus 4.8. Its independent SWE-Bench Verified score is on the vals.ai leaderboard and the ModelFit benchmark page. Cloud only. Parameter count undisclosed (cloud API). Source: https://www.vals.ai/benchmarks/swebench.
- [Claude Opus 4.8](https://modelfit.io/models/claude-opus-4.8): Cloud API, undisclosed size. Best for: Coding, Agentic, Reasoning. Anthropic's current flagship and the most capable model for Claude Code. Best-in-class agentic tool use and long-horizon coding. Parameter count undisclosed (cloud API). Cloud only. Source: https://www.anthropic.com/claude.
- [Claude Fable 5](https://modelfit.io/models/claude-fable-5/): Cloud API, undisclosed size. Best for: Coding, Agentic, Reasoning. Jun 9, 2026 Anthropic release: the first publicly available Mythos-class model, a tier above Opus 4.8. Anthropic reports it took the open coding lead on their own benchmarks (precise scores not published in primary-source text). Parameter count undisclosed (cloud API). Cloud only. Source: https://www.anthropic.com/claude.

### DeepSeek
- [DeepSeek V4 Pro](https://modelfit.io/models/deepseek-v4-pro/): Cloud API, 1600B params, 49B active MoE, open weights available. Best for: Frontier reasoning, Coding. Apr 24, 2026. 1.6T total, 49B active MoE. MIT license. Not realistically runnable on consumer Mac. API $0.87 per 1M output tokens (DeepSeek pricing page, Jul 2026). An August 0813 revision refreshed the API weights; same architecture, same canonical page. Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro.
- [DeepSeek-V3](https://modelfit.io/models/deepseek-v3/): Cloud API, 671B params, 37B active MoE, open weights available. Best for: Quality, Coding. Cloud/API only DeepSeek-V3 mixture of experts model with exceptional coding capabilities. Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.
- [DeepSeek-V3-0324](https://modelfit.io/models/deepseek-v3-0324/): Cloud API, 671B params, 37B active MoE, open weights available. Best for: Quality, Coding. Cloud/API only DeepSeek-V3-0324, March 2025 update with enhanced reasoning capabilities. Source: https://huggingface.co/deepseek-ai/DeepSeek-V3-0324.
- [DeepSeek-R1](https://modelfit.io/models/deepseek-r1-full/): Cloud API, 671B params, 37B active MoE, open weights available. Best for: Reasoning, Quality. Cloud/API only full DeepSeek-R1 671B MoE with advanced reasoning capabilities. Source: https://huggingface.co/deepseek-ai/DeepSeek-R1.
- [DeepSeek V4.1 Flash](https://modelfit.io/models/deepseek-v4.1-flash): Cloud API, 552B params, 16B active MoE, open weights available. Best for: Agentic coding, Reasoning, Long context, Vision. Sep 10, 2026 refresh of V4 Flash, MIT. Multimodal (image + text) MoE: 552B backbone plus a 196B sparsely-accessed Engram memory, 8B active on prefill / 16B on decode, 1M context. FP8 weights on Hugging Face (~390k downloads in week 1). No Ollama library tag; far beyond consumer hardware locally. Source: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash.
- [DeepSeek V4 Flash 0731](https://modelfit.io/models/deepseek-v4-flash/): Cloud API, 284B params, 13B active MoE, open weights available. Best for: Reasoning, Coding, Long context. Official 0731 release (Jul 31, 2026, MIT), superseding the April preview: same architecture, re-post-trained with far stronger agent skills. 284B total, 13B active MoE, 1M context. Community 2-bit dynamic quants fit 128GB Macs via MLX, llama.cpp and the ds4 engine. No Ollama library tag yet. Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731.

### Google
- [Gemini 2.5 Pro](https://modelfit.io/models/gemini-2.5-pro): Cloud API, undisclosed size. Best for: Quality, Analysis. Cloud/API only Gemini 2.5 Pro with advanced reasoning and multimodal capabilities. Parameter count undisclosed (cloud API). Source: https://deepmind.google/technologies/gemini/.
- [Gemini 2.5 Flash](https://modelfit.io/models/gemini-2.5-flash): Cloud API, undisclosed size. Best for: Speed, Chat. Cloud/API only Gemini 2.5 Flash optimized for speed and efficiency. Parameter count undisclosed (cloud API). Source: https://deepmind.google/technologies/gemini/.
- [Gemini 3.1 Pro](https://modelfit.io/models/gemini-3.1-pro): Cloud API, undisclosed size. Best for: Multimodal, Long context, Quality. Apr 2026 Google flagship. Strong multimodal, 2M context. Cloud only. Parameter count undisclosed (cloud API). Source: https://deepmind.google/technologies/gemini/.

### Inkling
- [Inkling Small](https://modelfit.io/models/inkling-small/): Cloud API, 276B params, 12B active MoE, open weights available. Best for: Agentic coding, Multimodal, Reasoning. Thinking Machines Lab open-weight MoE (Jul 2026, Apache 2.0): 276B total, 12B active, routed to 6 of 256 experts plus 2 shared. Text, image and audio input. SWE-bench Verified 80.2% per its model card. No Ollama tag and mainline llama.cpp does not load it yet; on a Mac the working path is MLX (the mlx-community 4-bit build is about 153 GB, a 192GB-class machine). Unsloth GGUFs are published for when llama.cpp support lands. Source: https://huggingface.co/thinkingmachines/Inkling-Small.

### Kimi
- [Kimi K3](https://modelfit.io/models/kimi-k3/): Cloud API, 2800B params, 104B active MoE, open weights available. Best for: Agents, Multimodal, Long context. Jul 2026 open-weight multimodal agentic MoE from Moonshot (weights on HuggingFace). Official checkpoint is far beyond any consumer machine; ollama run kimi-k3:cloud streams it from Ollama Cloud rather than running locally. Source: https://huggingface.co/moonshotai/Kimi-K3.
- [Kimi K2 Instruct](https://modelfit.io/models/kimi-k2-instruct/): Cloud API, 1000B params, 32B active MoE, open weights available. Best for: Quality, Coding. 1T parameter MoE model (32B active). Cloud/API only, requires ~560GB RAM at Q4, not feasible locally even on Mac Studio M3 Ultra 512GB. Source: https://huggingface.co/moonshotai/Kimi-K2-Instruct.
- [Kimi K2.6](https://modelfit.io/models/kimi-k2.6): Cloud API, 1000B params, 32B active MoE, open weights available. Best for: Coding, Agentic tasks. Apr 20, 2026 Moonshot AI release. 1T total, 32B active MoE. SWE-Bench Pro 58.6%. GGUF available (Unsloth UD-Q4: ~250GB). Open weights, Mac Studio 192-512GB territory. Source: https://huggingface.co/moonshotai/Kimi-K2.6.
- [Kimi K2.7-Code](https://modelfit.io/models/kimi-k2.7-code): Cloud API, 1000B params, 32B active MoE, open weights available. Best for: Coding, Agentic tasks. Jun 12, 2026 Moonshot AI release. 1T total / 32B active MoE built on K2.6. Moonshot reports it cuts thinking-token usage ~30%, a self-reported figure. 256K context, Modified MIT license. Open weights, but 1T scale means cloud/API for nearly all users; Ollama exposes a cloud-only tag. Source: https://huggingface.co/moonshotai/Kimi-K2.7-Code.

### MiMo
- [Xiaomi MiMo-V2.6-Pro](https://modelfit.io/models/mimo-v2.6-pro): Cloud API, 1020B params, 42B active MoE, open weights available. Best for: Frontier agentic coding, Multimodal. Sept 21, 2026 Xiaomi flagship (MIT, XiaomiMiMo/MiMo-V2.6-Pro-RL on HF). Sparse MoE with 1.02T total / 42B activated parameters, omnimodal input and 1M tokens of context. Open weights, but beyond any consumer machine: API or a multi-GPU server. Source: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL.
- [Xiaomi MiMo-V2-Flash](https://modelfit.io/models/mimo-v2-flash/): Cloud API, 309B params, 15B active MoE, open weights available. Best for: Reasoning, Coding, Agentic tasks. Dec 16, 2025 Xiaomi release (MIT, XiaomiMiMo/MiMo-V2-Flash on HF). 309B total / 15B active MoE with hybrid attention and Multi-Token Prediction, 256K context. SWE-Bench Verified 73.4 and AIME 2025 94.1 per Xiaomi's model card. No Ollama build, but community GGUFs exist (bartowski, unsloth), so a ~Q4 build fits a 256GB-class machine via llama.cpp. Source: https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash.
- [Xiaomi MiMo-V2.6-Flash](https://modelfit.io/models/mimo-v2.6-flash): Cloud API, 309B params, 15B active MoE, open weights available. Best for: Agentic coding, Multimodal, Long context. Sept 21, 2026 Xiaomi release (MIT, XiaomiMiMo/MiMo-V2.6-Flash-RL on HF), successor to MiMo-V2-Flash. Sparse MoE with 309B total / 15B activated parameters, native text, image, video and audio input, 1M tokens of context. No Ollama build; community GGUFs exist, so a ~Q4 build fits a 256GB-class machine via llama.cpp. Source: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL.

### MiniMax
- [MiniMax M3](https://modelfit.io/models/minimax-m3/): Cloud API, 428B params, 23B active MoE, open weights available. Best for: Coding, Agentic, Long context. Jun 1, 2026 MiniMax release. Native-multimodal MoE (~428B total / ~23B active) with 1M context via MiniMax Sparse Attention. SWE-Bench Pro 59.0%, Terminal-Bench 2.1 66.0%. Open weights; cloud/API for nearly all users. Source: https://huggingface.co/MiniMaxAI/MiniMax-M3.

### Mistral
- [Mistral Medium 3.5](https://modelfit.io/models/mistral-medium-3.5): Cloud API, 128B params, open weights available. Best for: Coding, Quality, Tool calling. Apr 30, 2026. 128B dense. SWE-Bench Verified 77.6%. Merges Magistral + Devstral 2 + Pixtral. GGUF available but 96GB+ Mac required. Open weights: mistralai/Mistral-Medium-3.5-128B on Hugging Face (modified MIT). Source: https://huggingface.co/mistralai/Mistral-Medium-3.5-128B.

### Nemotron
- [NVIDIA Nemotron 3 Ultra](https://modelfit.io/models/nemotron-3-ultra/): Cloud API, 550B params, 55B active MoE, open weights available. Best for: Reasoning, Agentic, Long context. Jun 4, 2026 NVIDIA release. 550B total / 55B active hybrid Mamba-2 + MoE, up to 1M context. SWE-Bench Verified 70.7. Weights are open (OpenMDW-1.1, nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 on HF; community GGUF is ~188GB even at 2-bit), so self-hosting needs a ~256GB-class multi-node rig. No consumer tier fits it and Ollama's tag is cloud-only, hence listed as cloud. Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16.

### OpenAI
- [GPT-4o](https://modelfit.io/models/gpt-4o/): Cloud API, undisclosed size. Best for: Quality, Coding. Cloud/API only GPT-4 Omni model with multimodal capabilities. Parameter count undisclosed (cloud API). Source: https://platform.openai.com/docs/models.
- [GPT-4o mini](https://modelfit.io/models/gpt-4o-mini/): Cloud API, undisclosed size. Best for: Chat, Speed. Cloud/API only compact version of GPT-4o with faster responses. Parameter count undisclosed (cloud API). Source: https://platform.openai.com/docs/models.
- [GPT-5.5](https://modelfit.io/models/gpt-5.5): Cloud API, undisclosed size. Best for: Quality, Reasoning, Multimodal. Apr 23, 2026 OpenAI flagship. Pro and Instant variants follow May 5. Cloud only. Parameter count undisclosed (cloud API). Source: https://platform.openai.com/docs/models.

### Qwen
- [Qwen3.8 2.4T A95B](https://modelfit.io/models/qwen3.8-2.4t-a95b): Cloud API, 2400B params, 95B active MoE, open weights available. Best for: Frontier reasoning, Agentic coding. Aug 12, 2026 Qwen flagship cloud MoE. 2.4T total parameters, 95B active (512 experts, 10 routed per token), BF16 weights on Hugging Face under the Qwen3.8-Max license (not Apache). Always-on thinking with adjustable reasoning effort. API-only in practice - no consumer hardware comes close. Source: https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B.
- [Qwen3.7-Plus](https://modelfit.io/models/qwen3.7-plus): Cloud API, undisclosed size. Best for: Multimodal, Agentic, Vision. Jun 1, 2026 Alibaba Qwen release. Proprietary multimodal agent model (vision + GUI control + code) on a 1M-token context, via Alibaba Cloud Model Studio / Bailian API. API-only, no open weights. Parameter count undisclosed. Source: https://qwen.ai.

### Zhipu
- [GLM-5.2](https://modelfit.io/models/glm-5.2): Cloud API, 753B params, open weights available. Best for: Coding, Reasoning, Agentic. Jun 13, 2026 coding-first flagship from Z.ai; MIT weights published on HuggingFace in July 2026 (753B total MoE). Works with Claude Code, Cline and OpenCode via the API. Weights are far beyond consumer hardware; use the API or hosted providers. Source: https://huggingface.co/zai-org/GLM-5.2.
- [GLM-5.3](https://modelfit.io/models/glm-5.3): Cloud API, 753B params, open weights available. Best for: Coding, Reasoning, Agentic. Z.ai flagship refresh (Aug 25, 2026), 753B total MoE with 1M context, MIT weights on Hugging Face. Successor to GLM-5.2 for Claude Code / OpenCode-style agentic use via API. Far beyond consumer hardware locally. Source: https://huggingface.co/zai-org/GLM-5.3.
- [GLM-5](https://modelfit.io/models/glm-5/): Cloud API, 744B params, 40B active MoE, open weights available. Best for: Reasoning, Quality. Cloud/API only GLM-5, Zhipu AI's latest reasoning model with MoE architecture. Source: https://huggingface.co/zai-org/GLM-5.
- [GLM-5.1](https://modelfit.io/models/glm-5.1): Cloud API, 744B params, 40B active MoE, open weights available. Best for: Coding, Reasoning, Long context. Apr 7, 2026 from Z.ai. 754B total, 40B active MoE with DeepSeek Sparse Attention. SWE-Bench Pro 58.4. MIT license. Local GGUF available (200-220GB). Source: https://huggingface.co/zai-org/GLM-5.1.
- [GLM-5.3 Flash](https://modelfit.io/models/glm-5.3-flash): Cloud API, 321B params, 18B active MoE, open weights available. Best for: Chat, Coding. Z.ai fast tier (Aug 25, 2026), 321B total / 18B active MoE, natively multimodal, MIT weights on Hugging Face. Mainline llama.cpp supports it (glm5-next architecture): the unsloth UD-IQ4_XS GGUF is about 157 GB, so a 192GB-class machine can run it locally; most users will reach it through the API. Source: https://huggingface.co/zai-org/GLM-5.3-Flash.
- [GLM-4 Plus](https://modelfit.io/models/glm-4-plus/): Cloud API, undisclosed size. Best for: Quality, Analysis. Cloud/API only GLM-4 Plus, Zhipu AI's high-performance general purpose model. Parameter count undisclosed (cloud API). Source: https://z.ai/.

### xAI
- [Grok 4.3](https://modelfit.io/models/grok-4.3): Cloud API, undisclosed size. Best for: Reasoning, Real-time data, Coding. Apr 17, 2026 xAI flagship. Full rollout May 6. 1M context (Apr 30 expansion). Cloud only. Parameter count undisclosed (cloud API). Source: https://x.ai/grok.