NVIDIA GeForce RTX 4070
12 GB · 504.2 GB/s · 200 W
In short
The NVIDIA GeForce RTX 4070 has 12 GB of GDDR6X memory with 504.2 GB/s of bandwidth, which sets how fast it generates text. At 128k context it runs 29 of the 125 open models we track fully in memory; the most capable is Gemma 4 26B-A4B at Q5_K_M, at about 32 tok/s (estimated). We have no current price for the NVIDIA GeForce RTX 4070 in the US.
Updated . Speeds are estimates; how they are worked out.
Prices
Every retailer we track in the US, cheapest first. Buy links may earn us a commission.
No retailer in this market lists it right now.
What two of them run
Every model in the catalogue, with the best quant for this hardware. Estimates, not measurements.
| Model | Quant | Fits | Memory | Max context | Decode @128k | Prefill | Good for |
|---|---|---|---|---|---|---|---|
| Gemma 4 26B-A4B | Q5_K_M | fits | 22.5 GB | 138k221k with q8 cache | 32tok/sest. | 951 tok/s | chat, coding, agents, long documents |
| gpt-oss 20B OpenAI | MXFP4 | fits | 16.1 GB | 128k | 33tok/sest. | 616 tok/s | chat, coding, agents, long documents |
| Gemma 4 12B | Q8_0 | fits | 16.0 GB | 256k | 21tok/sest. | 395 tok/s | chat, coding, long documents |
| Qwen 3.5 9B Alibaba Qwen | Q8_0 | fits | 14.8 GB | 256k | 23tok/sest. | 470 tok/s | chat, coding, long documents |
| Qwen 3.5 4B Alibaba Qwen | Q8_0 | fits | 10.1 GB | 256k | 35tok/sest. | 732 tok/s | chat, coding, long documents |
| Gemma 4 E4B | Q8_0 | fits | 13.0 GB | 128k | 26tok/sest. | 659 tok/s | chat, coding, long documents |
| LFM2.5 8B-A1B Liquid AI | Q8_0 | fits | 16.2 GB | 125k | 25tok/sest. | 881 tok/s | chat, coding, long documents |
| Llama 3.1 8B Meta | Q4_K_M | fits | 22.4 GB | 128k | 14tok/sest. | 260 tok/s | chat, coding |
| Qwen2.5 0.5B Alibaba Qwen | Q8_0 | fits | 3.8 GB | 32k | 146tok/sest. | 1980 tok/s | chat, coding, long documents |
| Qwen3.5 0.8B Alibaba Qwen | Q8_0 | fits | 4.1 GB | 256k | 128tok/sest. | 2640 tok/s | chat, coding, agents, long documents |
| Qwen3.5 2B Alibaba Qwen | Q5_K_M | fits | 4.7 GB | 256k | 102tok/sest. | 1688 tok/s | chat, coding, agents, long documents |
| DeepSeek R1 Distill Qwen 1.5B DeepSeek | Q8_0 | fits | 7.1 GB | 128k | 55tok/sest. | 885 tok/s | chat, coding, long documents |
| Mellum2.1 12B A2.5B Thinking JetBrains | Q8_0 | fits | 15.7 GB | 128k | 41tok/sest. | 1019 tok/s | chat, coding, agents, long documents |
| Spark X2.5 4B XHToken | mlx-4bit | fits | 9.2 GB | 458k767k with q8 cache | 39tok/sest. | 724 tok/s | chat, coding, agents, long documents |
| MiniCPM5 2B openbmb | Q8_0 | fits | 9.6 GB | 128k | 38tok/sest. | 476 tok/s | chat, coding, agents, long documents |
| FrogNano 4B 2609 Microsoft | Q8_0 | fits | 10.2 GB | 256k | 35tok/sest. | 732 tok/s | chat, coding, agents, long documents |
| LFM2.5 2.6B Liquid AI | Q8_0 | fits | 12.0 GB | 128k | 29tok/sest. | 610 tok/s | chat, coding, agents, long documents |
| LFM2.5 VL 3B Liquid AI | Q8_0 | fits | 12.0 GB | 32k | 29tok/sest. | 587 tok/s | chat, coding, agents, long documents |
| Kimi VL A3B Thinking 2506 Moonshot AI | Q8_0 | fits | 21.5 GB | 128k | 26tok/sest. | 640 tok/s | chat, coding, long documents |
| Kimi VL A3B Moonshot AI | mlx-8bit | fits | 22.2 GB | 128k | 26tok/sest. | 640 tok/s | chat, coding, agents, long documents |
| GLM 4 9B 0414 Z.ai | mlx-5bit | fits | 14.5 GB | 32k | 23tok/sest. | 212 tok/s | chat, coding |
| granite vision 4.1 4b IBM | Q8_0 | fits | 15.2 GB | 128k | 22tok/sest. | 377 tok/s | chat, coding, long documents |
| granite 4.2 3b IBM | Q8_0 | fits | 15.5 GB | 128k | 21tok/sest. | 384 tok/s | chat, coding, agents, long documents |
| GLM 4.6V Flash Z.ai | Q8_0 | fits | 16.2 GB | 128k | 20tok/sest. | 206 tok/s | chat, coding |
| DeepSeek R1 Distill Qwen 7B DeepSeek | Q8_0 | fits | 16.4 GB | 128k | 20tok/sest. | 316 tok/s | chat, coding, long documents |
| GLM 4.1V 9B Thinking Z.ai | mlx-8bit | fits | 17.8 GB | 64k | 18tok/sest. | 206 tok/s | chat, coding |
| Ministral 3 3B Instruct 2512 Mistral AI | Q8_0 | fits | 18.3 GB | 170k256k with q8 cache | 18tok/sest. | 368 tok/s | chat, coding, agents, long documents |
| Muse Glimmer 30B meta-models | Q5_K_M | fits | 22.3 GB | 128k | 14tok/sest. | 175 tok/s | chat |
| DeepSeek R1 Distill Llama 8B DeepSeek | Q4_K_M | fits | 22.4 GB | 128k | 14tok/sest. | 260 tok/s | chat, coding |
| Qwen 3.8 27B Alibaba Qwen | UD-Q4_K_M | offload | 25.3 GB | 91k159k with q8 cache | 8tok/sest. | 151 tok/s | chat |
| Qwen 3.5 122B-A10B Alibaba Qwen | UD-Q2_K_XL | offload | 44.0 GB | – | 8tok/sest. | 200 tok/s | chat |
| Qwen 3.6 27B Alibaba Qwen | Q4_K_M | offload | 25.7 GB | 86k150k with q8 cache | 8tok/sest. | 150 tok/s | chat |
| Gemma 4 31B | Q4_K_M | offload | 29.7 GB | 45k89k with q8 cache | 5tok/sest. | 130 tok/s | chat |
| Mistral Small 4 119B Mistral AI | UD-Q2_K_XL | offload | 42.1 GB | – | 11tok/sest. | 177 tok/s | chat |
| Qwen 3.6 35B-A3B Alibaba Qwen | MXFP4 | offload | 24.6 GB | 56k87k with q8 cache | 30tok/sest. | 746 tok/s | chat, coding, agents, long documents |
| Qwen 3.5 27B Alibaba Qwen | Q4_K_M | offload | 25.6 GB | 87k152k with q8 cache | 8tok/sest. | 150 tok/s | chat |
| Qwen 3.5 35B-A3B Alibaba Qwen | Q4_K_M | offload | 24.9 GB | 44k70k with q8 cache | 28tok/sest. | 741 tok/s | chat, coding, agents, long documents |
| GLM 4.7 Flash Z.ai | Q4_K_M | offload | 25.5 GB | 80k139k with q8 cache | 12tok/sest. | 344 tok/s | chat, coding, agents |
| Devstral Small 2 24B Mistral AI | Q4_K_M | offload | 35.2 GB | 51k93k with q8 cache | 3tok/sest. | 112 tok/s | chat |
| Nemotron 3.5 Lightning 30B-A3B NVIDIA | MXFP4 | offload | 24.3 GB | 3k4k with q8 cache | 47tok/sest. | 1013 tok/s | chat, coding, agents, long documents |
| Kimi Linear 48B-A3B Moonshot AI | Q4_K_M | offload | 30.9 GB | – | 27tok/sest. | 490 tok/s | chat, coding, long documents |
| Qwen3 Coder 30B-A3B Alibaba Qwen | Q4_K_M | offload | 31.1 GB | 44k79k with q8 cache | 5tok/sest. | 189 tok/s | chat, coding, agents |
| Gemma 3 27B | Q4_K_M | offload | 27.7 GB | 70k128k with q8 cache | 6tok/sest. | 175 tok/s | chat |
| Mistral Small 3.2 24B Mistral AI | Q4_K_M | offload | 35.2 GB | 51k93k with q8 cache | 3tok/sest. | 112 tok/s | chat |
| Ministral 3 14B Mistral AI | Q4_K_M | offload | 29.5 GB | 86k157k with q8 cache | 5tok/sest. | 158 tok/s | chat, coding |
| Qwen3 14B Alibaba Qwen | Q4_K_M | offload | 30.2 GB | 40k | 5tok/sest. | 133 tok/s | chat, coding |
| DeepSeek R1 0528 Qwen3 8B DeepSeek | Q4_K_M | offload | 24.5 GB | 116k128k with q8 cache | 9tok/sest. | 227 tok/s | chat, coding |
| Qwen3 8B Alibaba Qwen | Q4_K_M | offload | 24.5 GB | 40k | 9tok/sest. | 227 tok/s | chat, coding |
| Phi-4 14B Microsoft | Q4_K_M | offload | 35.1 GB | 16k | 3tok/sest. | 123 tok/s | chat, coding |
| Ministral 3 8B Mistral AI | Q4_K_M | offload | 23.7 GB | 121k220k with q8 cache | 11tok/sest. | 235 tok/s | chat, coding |
| Granite 4.2 8B IBM | Q4_K_M | offload | 27.0 GB | 102k128k with q8 cache | 6tok/sest. | 195 tok/s | chat, coding |
| Ministral 3 8B Reasoning 2512 Mistral AI | Q4_K_M | offload | 23.7 GB | 121k220k with q8 cache | 11tok/sest. | 235 tok/s | chat, coding |
| LFM2.5 24B A2B Liquid AI | Q4_K_M | offload | 26.1 GB | 88k128k with q8 cache | 9tok/sest. | 472 tok/s | chat, coding, agents |
| Qwen3.8 27B OBLITERATED OBLITERATUS | mlx-4bit | offload | 29.2 GB | 33k59k with q8 cache | 5tok/sest. | 139 tok/s | – |
| Ministral 3 14B Reasoning 2512 Mistral AI | Q4_K_M | offload | 29.5 GB | 86k157k with q8 cache | 5tok/sest. | 158 tok/s | chat, coding |
| DeepSeek R1 Distill Qwen 14B DeepSeek | Q4_K_M | offload | 34.2 GB | 68k126k with q8 cache | 3tok/sest. | 110 tok/s | chat, coding |
| Devstral Small 2507 Mistral AI | Q4_K_M | offload | 35.2 GB | 51k93k with q8 cache | 3tok/sest. | 113 tok/s | chat |
| Magistral Small 2506 Mistral AI | Q4_K_M | offload | 35.2 GB | 40k | 3tok/sest. | 113 tok/s | chat |
| Magistral Small 2507 Mistral AI | Q4_K_M | offload | 35.2 GB | 40k | 3tok/sest. | 113 tok/s | chat |
| Magistral Small 2509 Mistral AI | Q4_K_M | offload | 35.2 GB | 51k93k with q8 cache | 3tok/sest. | 112 tok/s | chat |
| DeepSeek V4 Flash DeepSeek | UD-Q2_K_XL | won't run | 102.8 GB | – | – | – | – |
| GLM 5.3 Flash Z.ai | UD-Q2_K_XL | won't run | 104.6 GB | – | – | – | – |
| Qwen 3.8 Flash Next Alibaba Qwen | UD-Q2_K_XL | won't run | 78.4 GB | – | – | – | – |
| MiniMax M2.7 MiniMax | UD-Q2_K_XL | won't run | 103.0 GB | – | – | – | – |
| Nemotron 3 Super 120B-A12B NVIDIA | UD-Q2_K_XL | won't run | 53.9 GB | – | – | – | – |
| Qwen3 Coder Next Alibaba Qwen | Q4_K_M | won't run | 50.1 GB | – | – | – | – |
| gpt-oss 120B OpenAI | Q4_K_M | won't run | 64.8 GB | – | – | – | – |
| GLM 4.5 Air Z.ai | Q2_K | won't run | 67.1 GB | – | – | – | – |
| Qwen3 Next 80B-A3B Alibaba Qwen | Q4_K_M | won't run | 50.1 GB | – | – | – | – |
| Qwen3 32B Alibaba Qwen | Q4_K_M | won't run | 52.3 GB | 12k23k with q8 cache | – | – | – |
| DeepSeek R1 Distill Qwen 32B DeepSeek | Q4_K_M | won't run | 52.3 GB | 12k23k with q8 cache | – | – | – |
| Llama 3.3 70B Meta | Q4_K_M | won't run | 81.5 GB | – | – | – | – |
| DeepSeek R1 DeepSeek | Q4_K_M | won't run | 387.1 GB | – | – | – | – |
| DeepSeek R1 0528 DeepSeek | Q4_K_M | won't run | 387.6 GB | – | – | – | – |
| DeepSeek R1 Distill Llama 70B DeepSeek | Q4_K_M | won't run | 81.5 GB | – | – | – | – |
| DeepSeek R1 Zero DeepSeek | Q4_K_M | won't run | 387.1 GB | – | – | – | – |
| DeepSeek V3 DeepSeek | Q4_K_M | won't run | 387.1 GB | – | – | – | – |
| DeepSeek V3 0324 DeepSeek | Q4_K_M | won't run | 387.1 GB | – | – | – | – |
| DeepSeek V3.1 DeepSeek | Q4_K_M | won't run | 388.0 GB | – | – | – | – |
| DeepSeek V3.1 Terminus DeepSeek | Q4_K_M | won't run | 388.0 GB | – | – | – | – |
| DeepSeek V3.2 DeepSeek | Q4_K_M | won't run | 388.0 GB | – | – | – | – |
| DeepSeek V4 Flash 0731 DeepSeek | MXFP4 | won't run | 158.2 GB | – | – | – | – |
| DeepSeek V4 Flash Vision Exp DeepSeek | MXFP4 | won't run | 167.0 GB | – | – | – | – |
| Devstral 2 123B Instruct 2512 Mistral AI | Q4_K_M | won't run | 115.6 GB | – | – | – | – |
| GLM 4.5 Z.ai | Q4_K_M | won't run | 249.4 GB | – | – | – | – |
| GLM 4.5V Z.ai | Q4_K_M | won't run | 84.1 GB | – | – | – | – |
| GLM 4.6 Z.ai | Q4_K_M | won't run | 248.6 GB | – | – | – | – |
| GLM 4.6V Z.ai | Q4_K_M | won't run | 90.5 GB | – | – | – | – |
| GLM 4.7 Z.ai | Q4_K_M | won't run | 249.4 GB | – | – | – | – |
| GLM 5 Z.ai | MXFP4 | won't run | 395.2 GB | – | – | – | – |
| GLM 5.1 Z.ai | MXFP4 | won't run | 432.8 GB | – | – | – | – |
| GLM 5.2 Z.ai | UD-Q4_K_M | won't run | 446.7 GB | – | – | – | – |
| GLM 5.3 Z.ai | Q8_0 | won't run | 759.1 GB | – | – | – | – |
| Kimi Dev 72B Moonshot AI | Q6_K | won't run | 101.8 GB | – | – | – | – |
| Kimi K2 Moonshot AI | Q4_K_M | won't run | 588.6 GB | – | – | – | – |
| Kimi K2 Instruct 0905 Moonshot AI | Q4_K_M | won't run | 589.0 GB | – | – | – | – |
| Kimi K2 Thinking Moonshot AI | Q4_K_M | won't run | 589.0 GB | – | – | – | – |
| Kimi K2.5 Moonshot AI | Q4_K_M | won't run | 589.0 GB | – | – | – | – |
| MiniMax M2 MiniMax | Q4_K_M | won't run | 161.7 GB | – | – | – | – |
| MiniMax M2.1 MiniMax | Q4_K_M | won't run | 161.7 GB | – | – | – | – |
| MiniMax M2.5 MiniMax | MXFP4 | won't run | 148.1 GB | – | – | – | – |
| MiniMax M3 MiniMax | MXFP4 | won't run | 256.5 GB | – | – | – | – |
| Mistral Medium 3.5 128B Mistral AI | Q4_K_M | won't run | 115.6 GB | – | – | – | – |
| granite 4.2 30b IBM | Q4_K_M | won't run | 50.6 GB | 19k35k with q8 cache | – | – | – |
Specifications
From the manufacturer's published specifications; sources are linked where we have them.
- GPU
- GeForce RTX 4070
- Memory
- 12 GB GDDR6X
- Memory bus
- 192-bit
- Memory bandwidth
- 504.2 GB/s
- FP16 compute
- 29.146 TFLOPS
- Software
- CUDA
- Board power
- 200 W
- PCIe generation
- Gen 4
- PCIe lanes
- x16
Compare with
Questions
- How much memory does the NVIDIA GeForce RTX 4070 have for AI models?
- The NVIDIA GeForce RTX 4070 has 12 GB of GDDR6X memory with 504.2 GB/s of bandwidth, which sets how fast it generates text.
- What is the best model the NVIDIA GeForce RTX 4070 can run locally?
- Gemma 4 26B-A4B (Q5_K_M) fits fully in its memory at 128k context and generates at about 32 tok/s, an estimate from its memory bandwidth. 29 of the 125 models we track fit at that context.
- Can the NVIDIA GeForce RTX 4070 run DeepSeek V4 Flash?
- No: even at UD-Q2_K_XL, DeepSeek V4 Flash needs about 103 GB at 128k context, more than the NVIDIA GeForce RTX 4070 and system RAM can hold.
- How much does the NVIDIA GeForce RTX 4070 cost in the US?
- We have no current price for the NVIDIA GeForce RTX 4070 in the US.
Speeds on this page are for 128k context. See the best hardware for local AI at every budget.