Sources confirmed 2026-10-03· ECI data 2026-10-21· LMArena 2026-10-20
Local LLM Leaderboard: What Your Machine Runs vs GPT-4 and GPT-5
A 16 GB MacBook Air (M5) runs Qwen3.5 9B, which scores 139.3 on the Epoch Capabilities Index: GPT-4.5 (Feb 2025) class, and above GPT-4 (Mar 2023) and GPT-4o (May 2024). Every local model ModelFit tracks, scored on one cross-era scale with the famous cloud models as reference lines.
- Models ranked
- 10842 with an ECI score
- Best on a 16 GB Mac
- Qwen3.5 9BECI 139.3, GPT-4.5 (Feb 2025) class
- Best open weights
- Kimi K3ECI 157.9, datacenter-class
- Top cloud model
- Claude Opus 5.5ECI 168.8, Anthropic
Key facts
A 16 GB MacBook Air (M5) runs Qwen3.5 9B, which scores 139.3 on the Epoch Capabilities Index: GPT-4.5 (Feb 2025) class, and above GPT-4 (Mar 2023) and GPT-4o (May 2024). The highest-scoring model that runs on a single machine is Qwen3.8 27B (ECI 151.7, needs 24 GB): GPT-5 (Aug 2025) class.
$ ollama run qwen3.5:9b
Scores are for full-precision or API runs, from Epoch AI and LMArena. ModelFit runs no benchmarks.
Your machine
What can your machine match?
Pick your Mac, iPhone or GPU. The ModelFit engine lists every model that fits, and this shows the one with the highest Epoch Capabilities Index score, next to the famous cloud model it matches.
Pick your machine
Your machine:
Best benchmarked model for MacBook Air M5 16 GB
Qwen3.5 9B
- ECI
- 139.3 (range 136.1 to 140.8)
- Class
- GPT-4.5 (Feb 2025) class
- Next step up
- o1 (Dec 2024), ECI 141.7
ollama run qwen3.5:9b
Score of the full-precision model. The Q4 build you run may score lower.
Get an email when a better model fits your MacBook Air M5 16 GB
You get one email when a new open-weight model becomes ModelFit's top local recommendation for this machine. The alert is not tied to any cloud model on this page.
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Fig. 1Epoch Capabilities Index
When did each Mac size catch GPT-4?
Each dot is an open-weight model at its release date, colored by the memory it needs. The lines are the famous cloud models: anything above a line scores higher than that model did.
- Minimum RAM
- 8 GB
- 16 GB
- 24 GB
- 32 GB
- 64 GB
- 128 GB
- 512 GB
- Rig (datacenter class)
In ModelFit's catalog, the earliest model that fits each memory size and reaches GPT-4 (Mar 2023) class or above:
- 16 GB: GPT-4 (Mar 2023) class since Apr 2025 (Qwen3 8B)
- 24 GB: GPT-4 (Mar 2023) class since Dec 2024 (Phi-4 14B)
- 32 GB: GPT-4 (Mar 2023) class since Dec 2024 (Phi-4 14B)
Data: Epoch AI, ECI (CC BY 4.0)
Table 1Capability leaderboard
Every local model, ranked
ECI 2026-10-21 · LMArena 2026-10-2042 of 42 models, 66 without an ECI score yet.
| Model | Class | Run | ||||
|---|---|---|---|---|---|---|
| Today's top cloud model: Claude Opus 5.5 · ECI 168.8 | ||||||
| Kimi K3datacenter-class | 157.9155.1 to 160.6 | Above GPT-5 (Aug 2025) | 1476 | 93.4% | Datacenter | |
| GLM-5.3datacenter-class | 156.0154.2 to 158.3 | Above GPT-5 (Aug 2025) | 1474 | 95.4% | Datacenter | |
| DeepSeek V4.1 Flashdatacenter-class | 155.9153.7 to 158.6 | Above GPT-5 (Aug 2025) | 1464 | Datacenter | ||
| DeepSeek V4 Flash 0731datacenter-class | 154.8152.1 to 156.9 | Above GPT-5 (Aug 2025) | 88.8% | Datacenter | ||
| GLM-5.3 Flashdatacenter-class | 153.1150.8 to 155.5 | Above GPT-5 (Aug 2025) | 1471 | 92.0% | Datacenter | |
| GLM-5.2datacenter-class | 151.9150.0 to 154.1 | GPT-5 (Aug 2025) class | 1470 | 82.8% | Datacenter | |
| Qwen3.8 27B | 151.7149.8 to 153.7 | GPT-5 (Aug 2025) class | 1440 | 86.0% | Q4 24 GBQ8 48 GB | $ qwen3.8:27b |
| Kimi K2.6datacenter-class | 150.9149.0 to 152.7 | GPT-5 (Aug 2025) class | 1455 | Datacenter | ||
| Inkling Smalldatacenter-class | 150.3147.2 to 152.1 | GPT-5 (Aug 2025) class | 1413 | 82.2% | Datacenter | |
| Kimi K2.7-Codedatacenter-class | 150.1148.2 to 151.8 | GPT-5 (Aug 2025) class | 78.2% | Datacenter | ||
| GPT-5 (Aug 2025) · ECI 150.0 | ||||||
| GLM-5.1datacenter-class | 149.8147.7 to 151.9 | Above Gemini 2.5 Pro (Jun 2025) | 1463 | Datacenter | ||
| DeepSeek V4 Prodatacenter-class | 149.1147.6 to 150.9 | Above Gemini 2.5 Pro (Jun 2025) | 1451 | 77.4% | Datacenter | |
| MiniMax M3datacenter-class | 146.9142.3 to 150.2 | Gemini 2.5 Pro (Jun 2025) class | 1433 | 75.0% | Datacenter | |
| Qwen3.6 27B | 146.2144.1 to 147.7 | Gemini 2.5 Pro (Jun 2025) class | 77.2% | Q4 32 GBQ8 48 GB | $ qwen3.6:27b | |
| NVIDIA Nemotron 3 Ultradatacenter-class | 146.0143.7 to 148.2 | Gemini 2.5 Pro (Jun 2025) class | 1445 | 69.0% | Datacenter | |
| GLM-5datacenter-class | 145.8143.8 to 147.7 | Gemini 2.5 Pro (Jun 2025) class | 1446 | Datacenter | ||
| Gemini 2.5 Pro (Jun 2025) · ECI 145.1 | ||||||
| Qwen3.6 35B-A3B | 143.7141.1 to 145.8 | Claude Opus 4 (May 2025) class | 73.4% | Q4 32 GBQ8 64 GB | $ qwen3.6:35b-a3b | |
| Gemma 4 31B | 143.0140.3 to 145.8 | Claude Opus 4 (May 2025) class | 1443 | Q4 32 GBQ8 48 GB | $ gemma4:31b | |
| Claude Opus 4 (May 2025) · ECI 142.5 | ||||||
| Qwen3.5 35B-A3B | 142.5140.4 to 145.0 | o1 (Dec 2024) class | 1395 | Q4 32 GBQ8 64 GB | $ qwen3.5:35b-a3b | |
| Gemma 4 26B-A4B | 141.9138.9 to 143.4 | o1 (Dec 2024) class | 1434 | Q4 24 GBQ8 48 GB | $ gemma4:26b | |
| o1 (Dec 2024) · ECI 141.7 | ||||||
| Mistral Medium 3.5datacenter-class | 141.4138.5 to 144.3 | GPT-4.5 (Feb 2025) class | 1421 | 66.4% | Datacenter | |
| Kimi K2 Instructdatacenter-class | 140.0136.3 to 141.9 | GPT-4.5 (Feb 2025) class | 1371 | Datacenter | ||
| GPT-OSS 120B | 139.8135.6 to 142.1 | GPT-4.5 (Feb 2025) class | 1366 | MXFP4 96 GB | $ gpt-oss:120b | |
| Qwen3.5 9B | 139.3136.1 to 140.8 | GPT-4.5 (Feb 2025) class | Q4 12 GBQ8 16 GB | $ qwen3.5:9b | ||
| Qwen3 235B A22B | 139.2134.6 to 140.5 | GPT-4.5 (Feb 2025) class | 1366 | Q4 192 GB | $ qwen3:235b-a22b-q4_K_M | |
| DeepSeek-R1 | 138.8136.0 to 140.2 | GPT-4.5 (Feb 2025) class | 1373 | Q4 512 GB | $ deepseek-r1:671b-q4_K_M | |
| Qwen3 14B | 138.2133.2 to 140.0 | GPT-4.5 (Feb 2025) class | Q4 16 GBQ8 24 GB | $ qwen3:14b-q4_K_M | ||
| GPT-OSS 20B | 137.7132.2 to 139.6 | GPT-4.5 (Feb 2025) class | 1288 | MXFP4 24 GB | $ gpt-oss:20b | |
| GPT-4.5 (Feb 2025) · ECI 136.7 | ||||||
| Qwen3 30B-A3B | 136.1130.9 to 138.7 | Above Claude 3.5 Sonnet (Jun 2024) | 1317 | Q4 32 GBQ8 48 GB | $ qwen3:30b-a3b-q4_K_M | |
| Qwen3 8B | 136.1129.9 to 138.3 | Claude 3.5 Sonnet (Jun 2024) class | Q4 12 GBQ8 16 GB | $ qwen3:8b-q4_K_M | ||
| DeepSeek-V3-0324datacenter-class | 135.8132.2 to 137.8 | Above Claude 3.5 Sonnet (Jun 2024) | 1375 | Datacenter | ||
| Llama 4 Maverick | 132.2127.7 to 134.3 | Claude 3.5 Sonnet (Jun 2024) class | 1287 | Q4 320 GB | $ llama4:maverick | |
| Phi-4 14B | 130.5124.7 to 132.0 | Claude 3.5 Sonnet (Jun 2024) class | 1217 | Q4 24 GBQ8 24 GB | $ phi4:14b-q4_K_M | |
| Claude 3.5 Sonnet (Jun 2024) · ECI 130.0 | ||||||
| Gemma 3 27B | 130.0124.5 to 132.0 | GPT-4o (May 2024) class | 1358 | Q4 32 GB | $ gemma3:27b | |
| Llama 4 Scout | 129.6125.1 to 131.2 | GPT-4o (May 2024) class | 1279 | Q4 96 GB | $ llama4:scout | |
| GPT-4o (May 2024) · ECI 129.0 | ||||||
| Mistral Small 3.1 | 127.5122.3 to 129.3 | GPT-4 Turbo (Apr 2024) class | 1277 | Q4 24 GBQ8 36 GB | $ mistral-small3.1:24b | |
| Llama 3.3 70B | 127.3122.5 to 129.4 | GPT-4 Turbo (Apr 2024) class | 1274 | Q4 64 GBQ8 96 GBQ6 96 GB | $ llama3.3:70b-instruct-q4_K_M | |
| GPT-4 Turbo (Apr 2024) · ECI 127.3 | ||||||
| Claude 3 Opus (Feb 2024) · ECI 126.9 | ||||||
| Llama 3.1 70B | 125.9120.6 to 128.2 | GPT-4 (Mar 2023) class | 1261 | Q4 64 GB | $ llama3.1:70b-instruct-q4_K_M | |
| GPT-4 (Mar 2023) · ECI 125.7 | ||||||
| Gemma 3 12B | 123.5114.2 to 128.8 | GPT-3.5 Turbo (Nov 2023) class | 1334 | Q4 16 GB | $ gemma3:12b | |
| Qwen2.5 7B | 118.5109.6 to 121.7 | GPT-3.5 Turbo (Nov 2023) class | Q4 8 GB | $ qwen2.5:7b-instruct-q4_K_M | ||
| GPT-3.5 Turbo (Nov 2023) · ECI 118.5 | ||||||
| Llama 3.1 8B | 116.5105.9 to 121.6 | 1187 | Q4 12 GBQ5 12 GBQ8 16 GB | $ llama3.1:8b-instruct-q4_K_M | ||
| Gemma 3 4B | 115.997.4 to 123.6 | 1291 | Q4 6 GBQ8 8 GB | $ gemma3:4b | ||
No ECI score yet (66)
These models have no Epoch Capabilities Index score yet. Some have an LMArena rating, shown in its column.
| Model | Class | Run | ||||
|---|---|---|---|---|---|---|
| Xiaomi MiMo-V2.6-Prodatacenter-class | 1491 | Datacenter | ||||
| Xiaomi MiMo-V2.6-Flashdatacenter-class | 1453 | Datacenter | ||||
| Qwen3.5 122B-A10B | 1417 | Q4 96 GB | $ qwen3.5:122b-a10b | |||
| Qwen3.5 27B | 1409 | 75.0% | Q4 24 GBQ8 48 GB | $ qwen3.5:27b | ||
| Nemotron 3 Super 120B-A12B | 1378 | IQ4_XS 96 GBQ4 128 GB | $ nemotron-3-super:120b | |||
| Granite 4.2 30B | 1361 | Q4 24 GB | $ granite4.2:30b | |||
| GLM-4.7-Flash 30B-A3B | 1351 | Q4 32 GB | $ glm-4.7-flash | |||
| DeepSeek-V3datacenter-class | 1333 | Datacenter | ||||
| Nemotron 3.5 Lightning 30B-A3B | 1331 | IQ4_XS 32 GBQ4 36 GB | $ nemotron-3.5-lightning:30b | |||
| Granite 4.2 8B | 1319 | Q4 8 GB | $ granite4.2:8b | |||
| Olmo 3.1 32B Instruct | 1311 | Q4 32 GB | $ olmo-3.1:32b | |||
| Granite 4.2 3B | 1294 | Q4 4 GB | $ granite4.2:3b | |||
| Granite 4.1 8B | 1290 | Q4 8 GB | $ granite4.1:8b | |||
| Llama 3.1 405B | 1284 | Q4 320 GB | $ llama3.1:405b-instruct-q4_K_M | |||
| Gemma 2 27B | 1231 | Q4 32 GB | $ gemma2:27b-instruct-q4_K_M | |||
| Gemma 2 9B | 1207 | Q4 12 GB | $ gemma2:9b-instruct-q4_K_M | |||
| Gemma 2 2B | 1156 | Q4 4 GB | $ gemma2:2b-instruct-q4_K_M | |||
| Phi-3 Medium 14B | 1138 | Q4 16 GB | $ phi3:14b-medium-4k-instruct-q4_K_M | |||
| Mixtral 8x7B Instruct | 1132 | Q4 48 GB | $ mixtral:8x7b | |||
| Llama 3.2 3B | 1110 | Q4 4 GB | $ llama3.2:3b-instruct-q4_K_M | |||
| Llama 3.2 1B | 1055 | Q4 2 GB | $ llama3.2:1b-instruct-q4_K_M | |||
| Cohere North Mini Code | Q4 32 GB | $ north-mini-code-1.0:q4_K_M | ||||
| DeepSeek-R1 Distill Llama 70B | Q4 64 GB | $ deepseek-r1:70b | ||||
| DeepSeek-R1 Distill Qwen 14B | Q4 16 GBQ8 24 GB | $ deepseek-r1:14b | ||||
| DeepSeek-R1 Distill Qwen 7B | Q4 8 GBQ8 16 GB | $ deepseek-r1:7b | ||||
| Devstral 2 123B | Q4 96 GB | $ devstral-2:123b | ||||
| Devstral Small 2 24B | Q4 24 GB | $ devstral-small-2:24b | ||||
| Gemma 3 1B | Q4 2 GBQ8 4 GB | $ gemma3:1b | ||||
| Gemma 4 12B | Q4 12 GBQ8 24 GB | $ gemma4:12b | ||||
| Gemma 4 E2B | Q4 4 GBQ8 8 GB | $ gemma4:e2b | ||||
| Gemma 4 E4B | Q4 6 GBQ8 16 GB | $ gemma4:e4b | ||||
| Granite 4.1 3B | Q4 4 GB | $ granite4.1:3b | ||||
| Laguna S 2.1 | Q4 128 GB | $ laguna-s-2.1:q4_K_M | ||||
| Laguna XS 2.1 | Q4 32 GB | $ laguna-xs-2.1:q4_K_M | ||||
| LFM2 24B-A2B | Q4 24 GB | $ lfm2:24b-a2b | ||||
| LFM2.5 8B-A1B | Q4 8 GB | $ lfm2.5:8b-a1b-q4_K_M | ||||
| Ling-3.0-flash | IQ4_XS 96 GBQ4 128 GB | |||||
| MiMo-V2.6-Distill-Qwen-9B | Q4 16 GB | |||||
| MiniCPM-V 4.5 8B | Q4 12 GB | $ minicpm-v4.5:8b | ||||
| Ministral 3 14B | Q4 16 GB | $ ministral-3:14b | ||||
| Ministral 3 3B | Q4 6 GB | $ ministral-3:3b | ||||
| Ministral 3 8B | Q4 12 GB | $ ministral-3:8b | ||||
| Mistral 7B Instruct | Q4 8 GBQ5 8 GBQ8 16 GB | $ mistral:7b-instruct-v0.3-q4_K_M | ||||
| Mistral Nemo 12B | Q4 16 GBQ8 24 GB | $ mistral-nemo:12b | ||||
| Mistral Small 22B | Q4 32 GB | $ mistral-small:22b | ||||
| Muse Glimmer 30B | Q4 32 GB | $ muse-glimmer:30b | ||||
| NVIDIA Nemotron Cascade 2 30B-A3B | Q6 36 GB | $ nemotron-cascade-2:30b | ||||
| Ornith 1.0 35B | Q4 32 GB | $ ornith:35b | ||||
| Ornith 1.0 9B | Q4 8 GB | $ ornith:9b | ||||
| Phi-3 Mini 3.8B | Q4 6 GB | $ phi3:3.8b-mini-4k-instruct-q4_K_M | ||||
| Phi-4 Mini 3.8B | Q4 6 GBQ8 8 GB | $ phi4-mini:3.8b | ||||
| Poolside Laguna XS.2 | Q4 36 GB | $ laguna-xs.2:q4_K_M | ||||
| Qwen2.5 0.5B | Q4 2 GB | $ qwen2.5:0.5b-instruct-q4_K_M | ||||
| Qwen2.5 1.5B | Q4 4 GB | $ qwen2.5:1.5b-instruct-q4_K_M | ||||
| Qwen2.5 14B | Q4 16 GB | $ qwen2.5:14b-instruct-q4_K_M | ||||
| Qwen2.5 3B | Q4 4 GB | $ qwen2.5:3b-instruct-q4_K_M | ||||
| Qwen2.5 Coder 14B | Q4 16 GB | $ qwen2.5-coder:14b | ||||
| Qwen2.5 Coder 7B | Q4 8 GB | $ qwen2.5-coder:7b | ||||
| Qwen3-Next 80B-A3B | Q4 72 GBQ8 128 GB | $ qwen3-next:80b | ||||
| Qwen3.5 0.8B | Q4 2 GB | $ qwen3.5:0.8b | ||||
| Qwen3.5 2B | Q4 4 GB | $ qwen3.5:2b | ||||
| Qwen3.5 4B | Q4 6 GBQ8 8 GB | $ qwen3.5:4b | ||||
| Qwen3.8 2.4T A95Bdatacenter-class | Datacenter | |||||
| Qwen3.8-Flash-Next | IQ4_XS 128 GBQ3 128 GBQ4 192 GB | |||||
| SmolLM2 360M | Q4 1 GB | $ smollm2:360m | ||||
| Xiaomi MiMo-V2-Flashdatacenter-class | Datacenter |
Scores are for full-precision or API runs; the Q4 build you run locally may score lower. The Epoch Capabilities Index (Epoch AI) and the LMArena text leaderboard (overall) (LMArena) are separate sources and are never mixed or averaged.
Fig. 2SWE-Bench Verified
Which local model is best for coding?
The best open-weight LLM on SWE-Bench Verified is Qwen3.8-27B (86.0%), which runs on a 24GB Mac, 11.0 points behind Claude Opus 5 (97.0%). SWE-Bench Verified is the headline real-world software-engineering benchmark; scores are third-party, raw-confirmed against each model's primary source.
Claude Opus 597.0% Cloud
DeepSeek V4 Pro 081396.4% Local
GPT-5.6-sol96.2% Cloud
Grok 4.695.6% Cloud
GPT-5.6-terra95.4% Cloud
GLM-5.395.4% Local
Kimi K393.4% Local
GLM-5.3 Flash92.0% Local
Table 2Coding leaderboard
SWE-Bench Verified leaderboard
Sources confirmed 2026-10-03| # | Model | SWE-Bench Verified | Type | Runs on | Get it | Source | Verified |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5Cloud API | 97.0% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 2 | DeepSeek V4 Pro 08131.6T / 49B MoE | 96.4% | Local | Multi-node rig | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 3 | GPT-5.6-solCloud API | 96.2% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 4 | Grok 4.6Cloud API | 95.6% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 5 | GPT-5.6-terraCloud API | 95.4% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 6 | GLM-5.3753B MoE | 95.4% | Local | Multi-node or 512GB-class rig | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 7 | Claude Fable 5Cloud API | 95.0% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 8 | Kimi K32.8T / 104B MoE | 93.4% | Local | Multi-node rig | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 9 | GPT-5.6-lunaCloud API | 93.0% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 10 | GLM-5.3 Flash321B / 18B MoE | 92.0% | Local | 192GB-class machine (llama.cpp) | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 11 | DeepSeek V4 Flash 0731284B / 13B MoE | 88.8% | Local | 192GB-class machine (llama.cpp) | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 12 | Claude Opus 4.8Cloud API | 88.6% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 13 | Grok 4.5Cloud API | 86.6% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 14 | Qwen3.8-27B27B | 86.0% | Local | 24GB Mac | $ qwen3.8:27bregistry-verified | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 15 | Qwen3.8-MaxCloud API | 85.6% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 16 | GLM-5.2753B MoE | 82.8% | Local | Multi-node or 512GB-class rig | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 17 | GPT-5.5Cloud API | 82.6% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 18 | Inkling Small276B / 12B MoE | 82.2% | Local | 192GB Mac (MLX) | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 19 | Claude Opus 4.7Cloud API | 82.0% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 20 | Gemini 3.8 FlashCloud API | 80.0% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 21 | Claude Sonnet 5Cloud API | 79.6% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-10-03 |
| 22 | Gemini 3.1 ProCloud API | 78.8% | Cloud | Cloud API | API only | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 23 | Kimi K2.7-Code1T / 32B MoE | 78.2% | Local | Multi-node or 512GB-class rig | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 24 | DeepSeek V4 Pro1.6T / 49B MoE | 77.4% | Local | Multi-node rig | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 25 | Qwen3.6-27B27B | 77.2% | Local | 32GB Mac | $ qwen3.6:27bregistry-verified | Qwen3.6-27B HuggingFace model card | 2026-09-29 |
| 26 | MiniMax M3428B / 23B MoE | 75.0% | Local | 256GB-class workstation | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 27 | Qwen3.5-27B27B | 75.0% | Local | 24GB Mac | $ qwen3.5:27bregistry-verified | Qwen3.5-27B HuggingFace model card | 2026-09-29 |
| 28 | Qwen3.6-35B-A3B35B / 3B MoE | 73.4% | Local | 32GB Mac | $ qwen3.6:35b-a3bregistry-verified | Qwen3.6-35B-A3B HuggingFace model card | 2026-09-29 |
| 29 | NVIDIA Nemotron 3 Ultra550B / 55B MoE | 69.0% | Local | 256GB-class multi-node rig | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 30 | Mistral Medium 3.5128B dense | 66.4% | Local | 96GB+ Mac | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
| 31 | Devstral 2 (2512)Open-weight 24B | 62.8% | Local | 32GB Mac | n/a | vals.ai SWE-Bench Verified leaderboard | 2026-09-29 |
Third-party SWE-Bench Verified scores. Each % is re-confirmed against the linked primary source and each Ollama tag re-probed against the registry (HTTP 200) on each refresh. Closed-API parameter counts are undisclosed ("Cloud API").
Looking for speed instead of capability? See measured tokens/sec from readers' machines.
Real-world calibration
Have this hardware? Share a measured run.
Measure a fixed Ollama reference model on your machine. This separate, opt-in contribution does not change ModelFit's third-party SWE-Bench leaderboard; results appear on the measured-speed leaderboard after the next site update.
$ npx -y @wecko-ai/modelfit bench --submit --cleanup
Methodology
How this leaderboard stays honest
- Capability
- Epoch Capabilities Index by Epoch AI (raw data, data of 2026-10-21). It combines many benchmarks into one scale, so a 2023 cloud model and a 2026 open-weight model can be compared directly.
- Human preference
- LMArena text leaderboard (overall) by LMArena (raw data, snapshot of 2026-10-20). Shown in its own column for context only.
- Coding
- third-party SWE-Bench Verified results, each re-confirmed against its primary source.
- The tier rule
- "Above GPT-4o" means the model's whole ECI range sits above GPT-4o's range. "GPT-4o class" means its score is at or above GPT-4o's but the two ranges overlap. Every such claim uses ECI only, never LMArena, and the two sources are never mixed or averaged.
- Precision
- scores are for full-precision or API runs. The Q4 build you run locally may score lower, and ModelFit does not invent a number for that gap.
- Unscored models
- stay unscored: a model is matched to a source only when the source scored the same weights. No estimate fills a gap, and ModelFit's own engine ratings are never shown as benchmark scores.
- Fit
- "runs on" and minimum memory come from the ModelFit engine, the same one behind every recommendation on the site. ModelFit runs no benchmarks.
- License
- ECI and LMArena data are CC BY 4.0, credited to Epoch AI and LMArena; selected and mapped to ModelFit's catalog by ModelFit.
FAQ
Frequently asked questions
Can a local LLM match GPT-4?
Yes. 38 of the 42 scored open-weight models reach GPT-4 (Mar 2023) class or above on the Epoch Capabilities Index. The smallest is Qwen3.5 9B, which needs 12 GB of memory and scores 139.3 against 125.7 for GPT-4 (Mar 2023).
What is the best model a 16 GB Mac can run?
A 16 GB MacBook Air (M5) runs Qwen3.5 9B, which scores 139.3 on the Epoch Capabilities Index: GPT-4.5 (Feb 2025) class, and above GPT-4 (Mar 2023) and GPT-4o (May 2024). It is the highest-scoring model the ModelFit engine fits in 16 GB. Pull it with: ollama run qwen3.5:9b.
Are these scores measured by ModelFit?
No. Capability scores come from Epoch AI's Epoch Capabilities Index (data of 2026-10-21) and the LMArena text leaderboard (snapshot of 2026-10-20), both CC BY 4.0. Coding scores are third-party SWE-Bench Verified results. ModelFit maps them to its catalog and runs no benchmarks. Scores are for full-precision or API runs; the Q4 build you run locally may score lower.
Why do ECI and LMArena rank models differently?
The Epoch Capabilities Index combines many benchmarks into one capability scale, while LMArena rates models by human votes in blind side-by-side chats. They measure different things: o1 (Dec 2024) is ahead on ECI (141.7 vs 136.7), but GPT-4.5 (Feb 2025) is ahead on LMArena (1417 vs 1366). This page keeps them in separate columns and every "class" or "above" claim uses ECI only.
Want the full analysis?
Detailed comparisons, RAM costs on Apple Silicon, and what runs on your Mac.