LLM VRAM requirements
How much GPU memory each open model needs at 4 bits, with 4,096 tokens of context and 20% runtime headroom, and the cheapest card that holds it at the latest checked prices. Open a model to change precision, context and batch, or to size LoRA and full fine-tuning.
Work out a custom scenario in the LLM VRAM calculator, or check how much VRAM your own card has. Speech, image, video, OCR and embedding models are sized by their publishers' own figures.
74 language models in 10 families and 8 speech, image, video, OCR and embedding models · prices checked 21 November 2026
Llama
- Llama 3.1 405B203 GBat 4-bit, weights only406B parameters · 812 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 256 GB MI325X at $3.80/hr.
- Llama 4 Maverick 17B 128E201 GBat 4-bit, weights only402B parameters · 803 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 256 GB MI325X at $3.80/hr.
- Llama 4 Scout 17B 16E54.3 GBat 4-bit, weights only109B parameters · 217 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 80 GB A100 at $1.09/hr.
- Llama 3.1 70B35.3 GBat 4-bit, weights only71B parameters · 141 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 48 GB RTX A6000 at $0.35/hr.
- Llama 3.3 70B35.3 GBat 4-bit, weights only71B parameters · 141 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 48 GB RTX A6000 at $0.35/hr.
- DeepSeek R1 Distill Llama 70B43.9 GBat 4-bit71B parameters · 171 GB at 16-bit. Cheapest card that holds it: 48 GB RTX A6000 at $0.35/hr.
- Llama 3.1 8B4.0 GBat 4-bit, weights only8.0B parameters · 16 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- DeepSeek R1 Distill Llama 8B5.5 GBat 4-bit8.0B parameters · 20 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Llama 3.2 3B1.6 GBat 4-bit, weights only3.2B parameters · 6 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- Llama 3.2 1B0.6 GBat 4-bit, weights only1.2B parameters · 2 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
Qwen
- Qwen3 Coder 480B A35B289 GBat 4-bit480B parameters · 1154 GB at 16-bit. No single card holds it; cheapest node: 8× L40 at $10.00/hr.
- Qwen3 235B A22B142 GBat 4-bit235B parameters · 565 GB at 16-bit. Cheapest card that holds it: 192 GB MI300X at $2.59/hr.
- Qwen3 235B A22B 2507142 GBat 4-bit235B parameters · 565 GB at 16-bit. Cheapest card that holds it: 192 GB MI300X at $2.59/hr.
- Qwen3 Next 80B A3B40.7 GBat 4-bit, weights only81B parameters · 163 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 48 GB RTX A6000 at $0.35/hr.
- Qwen2.5 72B45.2 GBat 4-bit73B parameters · 176 GB at 16-bit. Cheapest card that holds it: 48 GB RTX A6000 at $0.35/hr.
- Qwen2.5 32B20.9 GBat 4-bit33B parameters · 80 GB at 16-bit. Cheapest card that holds it: 24 GB A30 at $0.35/hr.
- Qwen2.5 Coder 32B20.9 GBat 4-bit33B parameters · 80 GB at 16-bit. Cheapest card that holds it: 24 GB A30 at $0.35/hr.
- DeepSeek R1 Distill Qwen 32B20.9 GBat 4-bit33B parameters · 80 GB at 16-bit. Cheapest card that holds it: 24 GB A30 at $0.35/hr.
- QwQ 32B20.9 GBat 4-bit33B parameters · 80 GB at 16-bit. Cheapest card that holds it: 24 GB A30 at $0.35/hr.
- Qwen3 32B20.9 GBat 4-bit33B parameters · 80 GB at 16-bit. Cheapest card that holds it: 24 GB A30 at $0.35/hr.
- Qwen3 30B A3B18.8 GBat 4-bit31B parameters · 74 GB at 16-bit. Cheapest card that holds it: 24 GB A30 at $0.35/hr.
- Qwen3 30B A3B 250718.8 GBat 4-bit31B parameters · 74 GB at 16-bit. Cheapest card that holds it: 24 GB A30 at $0.35/hr.
- Qwen3 Coder 30B A3B18.8 GBat 4-bit31B parameters · 74 GB at 16-bit. Cheapest card that holds it: 24 GB A30 at $0.35/hr.
- Qwen2.5 14B9.8 GBat 4-bit15B parameters · 36 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen2.5 Coder 14B9.8 GBat 4-bit15B parameters · 36 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- DeepSeek R1 Distill Qwen 14B9.8 GBat 4-bit15B parameters · 36 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen3 14B9.7 GBat 4-bit15B parameters · 36 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen3 8B5.6 GBat 4-bit8.2B parameters · 20 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen2.5 7B4.9 GBat 4-bit7.6B parameters · 19 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen2.5 Coder 7B4.9 GBat 4-bit7.6B parameters · 19 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- DeepSeek R1 Distill Qwen 7B4.9 GBat 4-bit7.6B parameters · 19 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen3 4B3.1 GBat 4-bit4.0B parameters · 10 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen3 4B 25073.1 GBat 4-bit4.0B parameters · 10 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen2.5 3B2.0 GBat 4-bit3.1B parameters · 8 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen2.5 Coder 3B2.0 GBat 4-bit3.1B parameters · 8 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen3 1.7B1.8 GBat 4-bit2.0B parameters · 5 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- DeepSeek R1 Distill Qwen 1.5B1.2 GBat 4-bit1.8B parameters · 4 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen2.5 1.5B1.1 GBat 4-bit1.5B parameters · 4 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen2.5 Coder 1.5B1.1 GBat 4-bit1.5B parameters · 4 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen3 0.6B1.0 GBat 4-bit0.8B parameters · 2 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen2.5 0.5B0.4 GBat 4-bit0.5B parameters · 1 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Qwen2.5 Coder 0.5B0.4 GBat 4-bit0.5B parameters · 1 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
DeepSeek
- DeepSeek V3.1342 GBat 4-bit, weights only685B parameters · 1369 GB at 16-bit. No single card has room for the weights alone; cheapest node that does: 8× L40 at $10.00/hr.
- DeepSeek R1 0528342 GBat 4-bit, weights only685B parameters · 1369 GB at 16-bit. No single card has room for the weights alone; cheapest node that does: 8× L40 at $10.00/hr.
- DeepSeek V3342 GBat 4-bit, weights only684B parameters · 1369 GB at 16-bit. No single card has room for the weights alone; cheapest node that does: 8× L40 at $10.00/hr.
- DeepSeek R1342 GBat 4-bit, weights only684B parameters · 1369 GB at 16-bit. No single card has room for the weights alone; cheapest node that does: 8× L40 at $10.00/hr.
Mistral and Mixtral
- Mixtral 8x22B v0.185.5 GBat 4-bit141B parameters · 339 GB at 16-bit. Cheapest card that holds it: 96 GB RTX PRO 6000 at $2.19/hr.
- Mixtral 8x7B v0.128.7 GBat 4-bit47B parameters · 113 GB at 16-bit. Cheapest card that holds it: 48 GB RTX A6000 at $0.35/hr.
- Mistral Small 3.2 24B 250615.2 GBat 4-bit24B parameters · 58 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Mistral Small 24B 250114.9 GBat 4-bit24B parameters · 57 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Devstral Small 250714.9 GBat 4-bit24B parameters · 57 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Mistral Nemo 24078.2 GBat 4-bit12B parameters · 30 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Mistral 7B v0.35.0 GBat 4-bit7.2B parameters · 18 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
Gemma
- gemma 3 27b13.7 GBat 4-bit, weights only27B parameters · 55 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- gemma 2 27b13.6 GBat 4-bit, weights only27B parameters · 54 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- gemma 3 12b6.1 GBat 4-bit, weights only12B parameters · 24 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- gemma 2 9b4.6 GBat 4-bit, weights only9.2B parameters · 18 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- gemma 3 4b2.2 GBat 4-bit, weights only4.3B parameters · 9 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- gemma 2 2b1.3 GBat 4-bit, weights only2.6B parameters · 5 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- gemma 3 1b0.5 GBat 4-bit, weights only1.0B parameters · 2 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
Phi
- phi 49.8 GBat 4-bit15B parameters · 36 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Phi 4 reasoning9.8 GBat 4-bit15B parameters · 36 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Phi 4 reasoning plus9.8 GBat 4-bit15B parameters · 36 GB at 16-bit. Cheapest card that holds it: 16 GB Tesla V100 at $0.26/hr.
- Phi 3 medium 4k7.0 GBat 4-bit, weights only14B parameters · 28 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- Phi 4 mini1.9 GBat 4-bit, weights only3.8B parameters · 8 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
- Phi 3 mini 4k1.9 GBat 4-bit, weights only3.8B parameters · 8 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
GPT-OSS
- gpt-oss-120b58.5 GBat 4-bit, weights only117B parameters · 234 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 80 GB A100 at $1.09/hr.
- gpt-oss-20b10.5 GBat 4-bit, weights only21B parameters · 42 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
GLM
- GLM 4.5217 GBat 4-bit358B parameters · 862 GB at 16-bit. Cheapest card that holds it: 256 GB MI325X at $3.80/hr.
- GLM 4.6216 GBat 4-bit357B parameters · 858 GB at 16-bit. Cheapest card that holds it: 256 GB MI325X at $3.80/hr.
- GLM 4.5 Air67.2 GBat 4-bit110B parameters · 266 GB at 16-bit. Cheapest card that holds it: 80 GB A100 at $1.09/hr.
Kimi
- Kimi K2 0905513 GBat 4-bit, weights only1026B parameters · 2053 GB at 16-bit. No single card has room for the weights alone; cheapest node that does: 8× RTX PRO 6000 at $20.00/hr.
- Kimi K2513 GBat 4-bit, weights only1026B parameters · 2053 GB at 16-bit. No single card has room for the weights alone; cheapest node that does: 8× RTX PRO 6000 at $20.00/hr.
Other open models
- Ministral 8B 24104.0 GBat 4-bit, weights only8.0B parameters · 16 GB at 16-bit. Cheapest card with room for the weights alone (cache not modelled): 16 GB Tesla V100 at $0.26/hr.
Speech, image, video, OCR and embedding models
Sized by their publishers' own figures, not the language-model estimate above.
- Whisperabout 10 GBlarge (1,550M parameters)Speech recognition, OpenAI. Cheapest card with that much: 16 GB Tesla V100 at $0.26/hr.
- Stable Diffusion 3.5 Medium9.9 GBStable Diffusion 3.5 Medium (2.5B)Image generation, Stability AI. Cheapest card with that much: 16 GB Tesla V100 at $0.26/hr.
- Wan2.18.19 GBT2V-1.3BVideo generation, Wan-AI. Cheapest card with that much: 16 GB Tesla V100 at $0.26/hr.
- Wan2.2at least 80 GBT2V-A14BVideo generation, Wan-AI. Cheapest card with that much: 80 GB A100 at $1.09/hr.
- HunyuanVideo60 GBHunyuanVideo, 720pVideo generation, Tencent. Cheapest card with that much: 80 GB A100 at $1.09/hr.
- olmOCRat least 12 GBolmOCR-2-7B-1025OCR and PDF conversion, Ai2. Cheapest card with that much: 24 GB A30 at $0.35/hr.
- MinerUat least 8 GBMinerU 4.0 Standard or AdvancedOCR and document parsing, OpenDataLab. Cheapest card with that much: 16 GB Tesla V100 at $0.26/hr.
- GTE-Qwen2about 32 GBGTE-Qwen2-7B-instruct, served with InfinityText embeddings, Alibaba. Cheapest card with that much: 48 GB RTX A6000 at $0.35/hr.
No model matches. Try a family name such as Llama, Qwen or DeepSeek.
All model memory estimates
| Model | Parameters | 4-bit weights, GB | Cache estimate |
|---|---|---|---|
| Kimi K2 0905 | 1026.41 billion | 513.2 | Not available |
| Kimi K2 | 1026.41 billion | 513.2 | Not available |
| DeepSeek V3.1 | 684.53 billion | 342.3 | Not available |
| DeepSeek R1 0528 | 684.53 billion | 342.3 | Not available |
| DeepSeek V3 | 684.49 billion | 342.2 | Not available |
| DeepSeek R1 | 684.49 billion | 342.2 | Not available |
| Qwen3 Coder 480B A35B | 480.15 billion | 240.1 | Config supported |
| Llama 3.1 405B | 405.85 billion | 202.9 | Not available |
| Llama 4 Maverick 17B 128E | 401.58 billion | 200.8 | Not available |
| GLM 4.5 | 358.34 billion | 179.2 | Config supported |
| GLM 4.6 | 356.79 billion | 178.4 | Config supported |
| Qwen3 235B A22B | 235.09 billion | 117.5 | Config supported |
| Qwen3 235B A22B 2507 | 235.09 billion | 117.5 | Config supported |
| Mixtral 8x22B v0.1 | 140.63 billion | 70.3 | Config supported |
| gpt-oss-120b | 117.00 billion | 58.5 | Not available |
| GLM 4.5 Air | 110.47 billion | 55.2 | Config supported |
| Llama 4 Scout 17B 16E | 108.64 billion | 54.3 | Not available |
| Qwen3 Next 80B A3B | 81.32 billion | 40.7 | Not available |
| Qwen2.5 72B | 72.71 billion | 36.4 | Config supported |
| Llama 3.1 70B | 70.55 billion | 35.3 | Not available |
| Llama 3.3 70B | 70.55 billion | 35.3 | Not available |
| DeepSeek R1 Distill Llama 70B | 70.55 billion | 35.3 | Config supported |
| Mixtral 8x7B v0.1 | 46.70 billion | 23.4 | Config supported |
| Qwen2.5 32B | 32.76 billion | 16.4 | Config supported |
| Qwen2.5 Coder 32B | 32.76 billion | 16.4 | Config supported |
| DeepSeek R1 Distill Qwen 32B | 32.76 billion | 16.4 | Config supported |
| QwQ 32B | 32.76 billion | 16.4 | Config supported |
| Qwen3 32B | 32.76 billion | 16.4 | Config supported |
| Qwen3 30B A3B | 30.53 billion | 15.3 | Config supported |
| Qwen3 30B A3B 2507 | 30.53 billion | 15.3 | Config supported |
| Qwen3 Coder 30B A3B | 30.53 billion | 15.3 | Config supported |
| gemma 3 27b | 27.43 billion | 13.7 | Not available |
| gemma 2 27b | 27.23 billion | 13.6 | Not available |
| Mistral Small 3.2 24B 2506 | 24.01 billion | 12.0 | Config supported |
| Mistral Small 24B 2501 | 23.57 billion | 11.8 | Config supported |
| Devstral Small 2507 | 23.57 billion | 11.8 | Config supported |
| gpt-oss-20b | 21.00 billion | 10.5 | Not available |
| Qwen2.5 14B | 14.77 billion | 7.4 | Config supported |
| Qwen2.5 Coder 14B | 14.77 billion | 7.4 | Config supported |
| DeepSeek R1 Distill Qwen 14B | 14.77 billion | 7.4 | Config supported |
| Qwen3 14B | 14.77 billion | 7.4 | Config supported |
| phi 4 | 14.66 billion | 7.3 | Config supported |
| Phi 4 reasoning | 14.66 billion | 7.3 | Config supported |
| Phi 4 reasoning plus | 14.66 billion | 7.3 | Config supported |
| Phi 3 medium 4k | 13.96 billion | 7.0 | Not available |
| Mistral Nemo 2407 | 12.25 billion | 6.1 | Config supported |
| gemma 3 12b | 12.19 billion | 6.1 | Not available |
| gemma 2 9b | 9.24 billion | 4.6 | Not available |
| Qwen3 8B | 8.19 billion | 4.1 | Config supported |
| Llama 3.1 8B | 8.03 billion | 4.0 | Not available |
| DeepSeek R1 Distill Llama 8B | 8.03 billion | 4.0 | Config supported |
| Ministral 8B 2410 | 8.02 billion | 4.0 | Not available |
| Qwen2.5 7B | 7.62 billion | 3.8 | Config supported |
| Qwen2.5 Coder 7B | 7.62 billion | 3.8 | Config supported |
| DeepSeek R1 Distill Qwen 7B | 7.62 billion | 3.8 | Config supported |
| Mistral 7B v0.3 | 7.25 billion | 3.6 | Config supported |
| gemma 3 4b | 4.30 billion | 2.2 | Not available |
| Qwen3 4B | 4.02 billion | 2.0 | Config supported |
| Qwen3 4B 2507 | 4.02 billion | 2.0 | Config supported |
| Phi 4 mini | 3.84 billion | 1.9 | Not available |
| Phi 3 mini 4k | 3.82 billion | 1.9 | Not available |
| Llama 3.2 3B | 3.21 billion | 1.6 | Not available |
| Qwen2.5 3B | 3.09 billion | 1.5 | Config supported |
| Qwen2.5 Coder 3B | 3.09 billion | 1.5 | Config supported |
| gemma 2 2b | 2.61 billion | 1.3 | Not available |
| Qwen3 1.7B | 2.03 billion | 1.0 | Config supported |
| DeepSeek R1 Distill Qwen 1.5B | 1.78 billion | 0.9 | Config supported |
| Qwen2.5 1.5B | 1.54 billion | 0.8 | Config supported |
| Qwen2.5 Coder 1.5B | 1.54 billion | 0.8 | Config supported |
| Llama 3.2 1B | 1.24 billion | 0.6 | Not available |
| gemma 3 1b | 1.00 billion | 0.5 | Not available |
| Qwen3 0.6B | 0.75 billion | 0.4 | Config supported |
| Qwen2.5 0.5B | 0.49 billion | 0.2 | Config supported |
| Qwen2.5 Coder 0.5B | 0.49 billion | 0.2 | Config supported |
Estimate training cost · Compare LLM API pricing · GPU rental prices