Qwen3.5 9B Instruct
Local OKExcellentA · 35
Best for: Quality, Coding, Reasoning·Pop: 86/100
Perf: est. 78 tok/s, typical error 5% · first token ~0.3s
Fits in 12 GB VRAM with room to spare. Best for quality, coding, reasoning on RTX 4070 SUPER.
Best Local AI Models for RTX 4070 SUPER (12GB)
The RTX 4070 SUPER improves on the base 4070 with more CUDA cores. Token generation is bandwidth-bound and both cards have 504 GB/s, so decode speed is est. 73 tok/s, typical error 5% against est. 73 tok/s for the base 4070, with the same 12GB VRAM capacity for efficient 7B-9B models.
Quick answer
The best local LLM for the RTX 4070 SUPER is Qwen3.5 9B Instruct (est. 78 tok/s, typical error 5%) on its 12GB VRAM. It uses ~7GB of VRAM; the RTX 4070 SUPER handles up to 9B parameter models at Q4. A 14B model at Q4 runs at est. 44 tok/s, typical error 15%. (ModelFit, 2026)
Sizing rule: a Q4 model needs about 0.6 GB of VRAM per billion parameters, and ModelFit budgets 90% of this card's 12GB for weights, context, and KV-cache. The per-size table below uses that same budget. Other strong fits: Gemma 4 12B (12B, ~8GB) and Ornith 1.0 9B (9B, ~5.6GB). On this card Qwen3.5 9B Instruct runs at est. 78 tok/s, typical error 5%. What it will not run: a 20B MoE (3.6B active) model does not fit in VRAM alone, so it spills to system RAM over PCIe and slows sharply.
$ollama run qwen3.5:9b
- Top pick
- Qwen3.5 9B Instruct
- Decode speed
- est. 78 tok/s, typical error 5%
- VRAM needed
- ~7 GB
Speeds are ModelFit estimates from memory bandwidth and model size, not measured benchmarks. Typical error: the median gap between the ModelFit speed model and measured runs on hardware the model was not fitted on, for this backend and model size.
Cite this page: ModelFit, RTX 4070 SUPER 12GB Local LLM (2026): 7B-9B at est. 73 tok/s, https://modelfit.io/gpu/rtx-4070-super/, updated October 2026, CC BY 4.0.
Specs
VRAM
12 GB GDDR6X
Speed (8B Q4)
est. 73 tok/stypical error 5%
Bandwidth
504 GB/s
Architecture
Ada Lovelace
Max model size
Up to 9B parameter models
Compatibility
10 excellent, 0 workable
Table 1
RTX 4070 SUPER estimated tokens/sec by model size
Q4_K_M · ModelFit estimate| Model size | Est. speed | Fit on 12GB |
|---|---|---|
| 7B | est. 84 tok/stypical error 5% | Fits in VRAM |
| 14B | est. 44 tok/stypical error 15% | Fits in VRAM |
| 20B MoE (3.6B active) | est. 73 tok/stypical error 15% | CPU offload (slow) |
| 32B | est. 2 tok/stypical error 5% | CPU offload (slow) |
| 35B MoE (3B active) | est. 40 tok/stypical error 15% | CPU offload (slow) |
| 70B | est. 1 tok/stypical error 10% | CPU offload (slow) |
| 120B MoE (5.1B active) | est. 25 tok/stypical error 15% | CPU offload (slow) |
Fig. 1RTX 4070 SUPER: estimated speed by model size
Decode speed, generic Q4_K_M model of each size at 4K context, 504 GB/s, typical error 5-15%
Fits in 12 GBCPU offload (slow)
7B84
14B44
20B MoE (3.6B active)73
32B2
35B MoE (3B active)40
70B1
120B MoE (5.1B active)25
tok/s est.ModelFit estimates, not measured benchmarks: decode speed of a generic Q4_K_M model of each size at 4K context on the RTX 4070 SUPER's 504 GB/s bandwidth. MoE rows scale by active parameters (decode reads only the active experts), so a 35B MoE runs far faster than a dense 32B. "CPU offload" sizes exceed the 12GB VRAM; dense models slow to a crawl there, MoE models degrade less because hot experts stay GPU-resident. What typical error means.
Fig. 2
Context costs VRAM too.
Qwen3.5 9B Instruct loads ~7 GB of weights; at 16k context the KV cache adds ~0.5 GB (still fits the ~11 GB usable VRAM), and at 64k it adds ~2.0 GB (still fits).
Weights 7.0 GBKV cache · 16k context 0.5 GBHeadroom 3.3 GBRuntime reserve 1.2 GB
The model leaves about 3.3 GB of the usable VRAM budget free after weights and 16k context.
KV-cache figures assume an fp16 cache, the llama.cpp/Ollama default. Standard GQA models use a size-class estimate (8 KV heads x 128 head dim class); hybrid linear-attention models (Qwen3.5/3.6, Qwen3-Next) use the exact per-token cost from their published config, since only their sparse full-attention layers cache KV. A q8_0 KV cache roughly halves either figure. Estimates, not measurements.
Get told when a model beats Qwen3.5 9B Instruct on your RTX 4070 SUPER
One email when a new open-weight model is a better fit for 12 GB of VRAM, plus the Thursday weekly.
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.
Where to buy the RTX 4070 SUPER
Storage & accessories for your model library
- Internal NVMe SSD · 2TBA Gen4 M.2 drive keeps your whole GGUF and quant collection on fast local storage, loading models straight off NVMe.Check price on Amazon
- USB4 NVMe Enclosure40Gbps external storage fast enough to run models from. Pair it with an M.2 drive for a portable model vault.Check price on Amazon
ModelFit may earn a commission on purchases through these links, at no extra cost to you.
RTX 4070 SUPER VRAM for AI: what actually fits?
Same 12GB GDDR6X and 504 GB/s bandwidth as the base RTX 4070, but extra CUDA cores push compute throughput higher. Because token generation is bandwidth-bound, decode speed is est. 73 tok/s, typical error 5% vs est. 73 tok/s; the extra compute mainly helps prompt reading. VRAM usage is identical to other 12GB cards. Best for users who want peak speed with 7B-9B models and found a good deal on the SUPER variant.
Table 2
RTX 4070 SUPER vs similar GPUs
| Hardware | Memory | Speed | Bandwidth |
|---|---|---|---|
| RTX 4070 SUPER | 12 GB | est. 73 tok/stypical error 5% | 504 GB/s |
| RTX 4070 | 12 GB | est. 73 tok/stypical error 5% | 504 GB/s |
| RTX 5070 | 12 GB | est. 98 tok/stypical error 5% | 672 GB/s |
| RTX 4070 Ti SUPER | 16 GB | est. 94 tok/stypical error 5% | 672 GB/s |
Recommended models
registry-verified10 modelsGemma 4 12B
Local OKOKB · 26
Best for: Chat, Coding, Multimodal·Pop: 80/100
Perf: est. 52 tok/s, typical error 15% · first token ~0.4s
Fits in 12 GB VRAM with room to spare. Best for chat, coding, multimodal on RTX 4070 SUPER.
Ornith 1.0 9B
Local OKExcellentA · 48
Best for: Agentic coding on small machines·Pop: 76/100
Perf: est. 78 tok/s, typical error 5% · first token ~0.3s
Fits in 12 GB VRAM with room to spare. Best for agentic coding on small machines on RTX 4070 SUPER.
Mellum2.1 12B-A2.5B Thinking
Local OKOKB · 25
Best for: Agentic coding, Small machines·Pop: 76/100
Perf: est. 141 tok/s, typical error 15% · first token ~0.1s
Fits in 12 GB VRAM with room to spare. Best for agentic coding, small machines on RTX 4070 SUPER.
Gemma 4 E4B (Q8)
Local OKExcellentA · 31
Best for: On-device, Mobile, Chat·Pop: 82/100
Perf: est. 79 tok/s, typical error 5% · first token ~1.3s
Fits in 12 GB VRAM with room to spare. Best for on-device, mobile, chat on RTX 4070 SUPER.
LFM2.5 8B-A1B
Local OKExcellentA · 49
Best for: On-device agents, tool calling, multilingual chat·Pop: 72/100
Perf: est. 196 tok/s, typical error 15% · first token ~0.1s
Fits in 12 GB VRAM with room to spare. Best for on-device agents, tool calling, multilingual chat on RTX 4070 SUPER.
NVIDIA Nemotron 3.5 Bolt E4B
Local OKOKB · 27
Best for: Voice assistants, Tool calling, Chat·Pop: 75/100
Perf: est. 98 tok/s, typical error 5% · first token ~0.2s
Fits in 12 GB VRAM with room to spare. Best for voice assistants, tool calling, chat on RTX 4070 SUPER.
MiMo-V2.6-Distill-Qwen-9B
Local OKExcellentA · 46
Best for: Agentic coding, Small machines·Pop: 60/100
Perf: est. 76 tok/s, typical error 5% · first token ~0.3s
Fits in 12 GB VRAM with room to spare. Best for agentic coding, small machines on RTX 4070 SUPER.
Granite Switch 4.1 8B
Local OKExcellentA · 41
Best for: Chat, Decisions, Safety checks·Pop: 64/100
Perf: est. 57 tok/s, typical error 5% · first token ~0.3s
Fits in 12 GB VRAM with room to spare. Best for chat, decisions, safety checks on RTX 4070 SUPER.
Ministral 3 8B
Local OKExcellentA · 44
Best for: Chat, Coding, Vision·Pop: 74/100
Perf: est. 65 tok/s, typical error 5% · first token ~0.3s
Fits in 12 GB VRAM with room to spare. Best for chat, coding, vision on RTX 4070 SUPER.
Video and image models on the RTX 4070 SUPER
Runs · Stated: the linked source says a GPU with that much VRAM runs the model. Runs · Published: a dated run on the named card with that much VRAM. unknown: no source yet. Exceeds VRAM: the complete download is larger than 12 GB, so part of it has to sit in system RAM. Every VRAM tier: video · image.
Video
| Model | Fits | Quality | App |
|---|---|---|---|
| HunyuanVideo-1.5Tencent Hunyuan Community License | Runs · PublishedRTX 4070 · 12 GB VRAM · ComfyUI example workflow, 480p I2V distilled fp8, 81 frames · ComfyUI issue #10855 | #38 of 48 in text to video4,292 votessnapshot 2026-09-22 | ComfyUI |
| LTX-2.3LTX-2 Community License Agreement | Runs · PublishedRTX 4060 · 8 GB VRAM · ComfyUI GGUF workflow, 720p 5 s in about 190 s, 32 GB RAM · ComfyUI issue #14345 | No independent score yet | Lightricks/LTX-2 ltx-pipelines, ComfyUI |
| LTX-2.5LTX-2.x Community License Agreement | Runs · Stated12 GB VRAM · LTX Desktop on Windows, weight streaming, 40 GB RAM, 540p · LTX Desktop compatibility matrix (Lightricks) | No independent score yet | Lightricks/LTX-2 ltx-pipelines, ComfyUI |
| MiniMax H3MiniMax H3 Community License Agreement | Runs · Stated12 GB VRAM · Diffusers int8, streamed blocks, 960x544, about 75 GB system RAM · Diffusers docs: MiniMax-H3 memory | #8 of 48 in text to video11,417 votessnapshot 2026-09-22 | stable-diffusion.cpp, ComfyUI |
| Wan2.2 A14BApache License 2.0 | Runs · PublishedRTX 2060 · 6 GB VRAM · ComfyUI Q8 GGUF, 480x720, 81 frames, 7 to 10 min, 28 GB RAM plus a 60 GB page file · ComfyUI issue #12451 | #42 of 48 in text to video10,458 votessnapshot 2026-09-22 | ComfyUI |
| Wan2.2 TI2V 5BApache License 2.0 | Runs · Stated8 GB VRAM · ComfyUI native offloading · ComfyUI docs: Wan2.2Complete: Comfy-Org repackaged FP16 with FP8 scaled UMT518.1 GB download | No independent score yet | ComfyUI |
| Mochi 1 previewApache License 2.0 | unknownComplete: calcuis GGUF Q3_K_M with Q4_0 T5 and FP8 VAE7.7 GB download | #48 of 48 in text to video5,893 votessnapshot 2026-09-22 | Diffusers, ComfyUI |
Quality: blind human votes from Arena (arena.ai), CC BY 4.0, snapshot 2026-09-22. Scores are for the Arena checkpoint, not your local quantized build.
Image
| Model | Fits | Quality | App |
|---|---|---|---|
| FLUX 3 Image KleinFLUX Kommunity License v1.0 | Runs · Stated8 GB VRAM · official ComfyUI int8 files · FLUX 3 Image Klein README (Black Forest Labs) | No independent score yet | Diffusers |
| FLUX.1 [dev]FLUX.1 [dev] Non-Commercial License v1.1.1 | Runs · Stated6 GB VRAM · stable-diffusion.cpp with GGUF weights · stable-diffusion.cpp FLUX docs | #78 of 82 in text to image49,221 votessnapshot 2026-10-07 | Diffusers, ComfyUI |
| Krea 2 TurboKrea 2 Community License Agreement v.1 | Runs · PublishedRTX 4070 · 12 GB VRAM · ComfyUI fp8, dynamic VRAM paging, 8 steps · ComfyUI issue #15365 | No independent score yet | Diffusers, stable-diffusion.cpp |
| Qwen-Image-2.1Qwen Research License Agreement | Runs · PublishedRTX 5070 Ti Laptop · 12 GB VRAM · stable-diffusion.cpp Q6_K, offload to CPU, 1024x1024 · stable-diffusion.cpp issue #2118 | #19 of 82 in text to image11,441 votessnapshot 2026-10-07 | Diffusers, stable-diffusion.cpp |
| SDXL Base 1.0CreativeML Open RAIL++-M License | Runs · Stated8 GB VRAM · SDXL 1.0, no app named · SDXL 1.0 announcement (Stability AI) | No independent score yet | ComfyUI, stable-diffusion.cpp |
| Stable Diffusion 3.5 LargeStability AI Community License | Runs · Stated11 GB VRAM · FP8 TensorRT build on RTX 40 and 50 · Stability AI: SD3.5 TensorRT models | #81 of 82 in text to image23,396 votessnapshot 2026-10-07 | ComfyUI |
| Z-Image-TurboApache 2.0 | Runs · Stated4 GB VRAM · stable-diffusion.cpp with GGUF weights · stable-diffusion.cpp Z-Image docs | #59 of 82 in text to image24,808 votessnapshot 2026-10-07 | Diffusers, stable-diffusion.cpp |
| FLUX.2 [dev]FLUX Non-Commercial License | Exceeds VRAMComplete: Q4_K_M GGUF34.8 GB download | #38 of 82 in text to image77,988 votessnapshot 2026-10-07 | Diffusers, ComfyUI, stable-diffusion.cpp |
| FLUX.2 [klein]Apache 2.0 (4B), FLUX Non-Commercial License (9B) | Exceeds VRAMComplete: Klein 4B Q8_0 GGUF12.5 GB download | No independent score yet | Diffusers, stable-diffusion.cpp |
Quality: blind human votes from Arena (arena.ai), CC BY 4.0, snapshot 2026-10-07. Scores are for the Arena checkpoint, not your local quantized build.
Similar GPUs for local AI
Speeds are estimates for Qwen3 8B Q4_K_M, typical error 5%.
Compatible model families
- QwenAlibaba Cloud: Qwen3.8 27B local flagship (Apache 2.0, vision)
- LlamaMeta: Most popular open-weight model family
- DeepSeekDeepSeek AI: V4.1 Flash: multimodal 552B MoE, API only
- MistralMistral AI: Excellent performance-per-parameter ratio
- GemmaGoogle DeepMind: Gemma 4 E4B: 4.5B effective, runs on 8GB Macs
- PhiMicrosoft: Best quality-per-gigabyte at small sizes
RTX 4070 SUPER FAQ: common questions
How much VRAM does the RTX 4070 SUPER have for LLMs?
The RTX 4070 SUPER has 12GB GDDR6X VRAM, identical to the base RTX 4070. Both share 504 GB/s bandwidth. The SUPER variant adds more CUDA cores for faster compute, but decode speed is bandwidth-bound: est. 73 tok/s, typical error 5% vs est. 73 tok/s.
What size LLM can I run on an RTX 4070 SUPER?
Up to 9B parameter models at Q4 quantization. Same model capacity as the base 4070 and RTX 3060. Decode speed is est. 73 tok/s, typical error 5%, level with the base 4070 (est. 73 tok/s) and 46% faster than the 3060.
Is the RTX 4070 SUPER worth it over the base 4070 for AI?
At launch, both cards shared the same $599 MSRP. At current street pricing the SUPER carries a premium for no decode speed gain (est. 73 tok/s, typical error 5% vs est. 73 tok/s), a poor trade for AI workloads. Buy it only if the gap narrows to a rounding error, or if you also use the GPU for gaming.
RTX 4070 SUPER vs RTX 5070 for local LLMs?
The RTX 5070 is faster (est. 98 tok/s, typical error 5% vs est. 73 tok/s) and, at current pricing, clearly cheaper. Both have 12GB VRAM. The 5070 wins on both price and performance for AI workloads.
How fast is a 27B-class model on the RTX 4070 SUPER?
The RTX 4070 SUPER's 12GB of VRAM cannot fit a 32B model comfortably. The largest size class it fits is 14B, at est. 44 tok/s, typical error 15%. The largest individual model in the catalog that fits is Ministral 3 14B (14B).
Related guides & benchmarks
More local AI guides
- How Much RAM (or VRAM) Do You Need for a Local LLM?The model-size-to-memory matrix: what each VRAM and RAM tier actually runs.
- Best LLM for MacBook (Apple Silicon)Unified-memory Macs run bigger models per dollar. Picks by RAM tier, M1 to M5.
- Local AI image generator on MacCompare image models by RAM and app.
- Local image generation on MacImage models and apps by Mac RAM.
Browse all NVIDIA GPUs for AI
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
Want personalized recommendations?
Use our interactive wizard to compare models across Apple Silicon and NVIDIA GPUs.