What can I run with 40 GB of VRAM?
With 40 GB of VRAM (or unified memory) and an 8K-token conversation, 22 of the 30 language models in our calculator fit comfortably, and the largest is Mixtral 8x7B Instruct v0.1.
Language models (LLMs)
| Model | Best precision | Memory | Verdict |
|---|---|---|---|
| Mixtral 8x7B Instruct v0.1Mistral AI | 4-bit (Q4_K_M) | 27.5 GB | Fits |
| Recursive FlintRecursive | 4-bit (Q4_K_M) | 25.3 GB | Fits |
| Qwen3.6 35B-A3BAlibaba Qwen | 4-bit (Q4_K_M) | 22.5 GB | Fits |
| Qwen3.8-35B-A3B (Empero Distill)Empero | 4-bit (Q4_K_M) | 21.9 GB | Fits |
| Nanbeige4.3 33BNanbeige | 4-bit (Q4_K_M) | 22.5 GB | Fits |
| Qwen2.5 32B InstructAlibaba Qwen | 4-bit (Q4_K_M) | 22.0 GB | Fits |
| Qwen3 32BAlibaba Qwen | 4-bit (Q4_K_M) | 21.9 GB | Fits |
| Gemma 4 31BGoogle | 8-bit (Q8_0) | 32.6 GB* | Fits |
| Qwen3.8 27BAlibaba Qwen | 8-bit (Q8_0) | 29.7 GB | Fits |
| Gemma 4 26B A4BGoogle | 8-bit (Q8_0) | 26.9 GB* | Fits |
| LFM2.5-24B-A2BLiquid AI | 8-bit (Q8_0) | 25.7 GB | Fits |
| gpt-oss-20bOpenAI | 16-bit (FP16/BF16) | 14.0 GB | Fits |
| Qwen2.5 14B InstructAlibaba Qwen | 16-bit (FP16/BF16) | 31.2 GB | Fits |
| Qwen3 14BAlibaba Qwen | 16-bit (FP16/BF16) | 30.9 GB | Fits |
| DeepSeek-R1-0528-Qwen3-8BDeepSeek | 16-bit (FP16/BF16) | 17.6 GB | Fits |
| Llama 3.1 8B InstructMeta | 16-bit (FP16/BF16) | 17.1 GB | Fits |
| Nemotron 3.5 Bolt E4BNVIDIA | 16-bit (FP16/BF16) | 15.1 GB* | Fits |
| Qwen2.5 7B InstructAlibaba Qwen | 16-bit (FP16/BF16) | 15.7 GB | Fits |
| Mistral 7B Instruct v0.3Mistral AI | 16-bit (FP16/BF16) | 15.6 GB | Fits |
| Spark-X2.6 4BXHToken | 16-bit (FP16/BF16) | 8.5 GB | Fits |
| Llama 3.2 3B InstructMeta | 16-bit (FP16/BF16) | 7.4 GB | Fits |
| LFM3 1.7BLiquid AI | 16-bit (FP16/BF16) | 3.6 GB | Fits |
| Kimi K3Moonshot AI | needs 1,508.9 GB at 4-bit | 1,508.9 GB | Too big |
| Mistral Large 4 (Le Chonk)Mistral AI | needs 635.5 GB at 4-bit | 635.5 GB | Too big |
| DeepSeek-V4.1-Flash-1017DeepSeek | needs 500.7 GB at 4-bit | 500.7 GB | Too big |
| Qwen3.8 Flash NextAlibaba Qwen | needs 119.8 GB at 4-bit | 119.8 GB | Too big |
| Ornith-1.5-122BOrnith | needs 79.9 GB at 4-bit | 79.9 GB | Too big |
| gpt-oss-120bOpenAI | needs 63.1 GB at 4-bit | 63.1 GB | Too big |
| Llama 3.1 70B InstructMeta | needs 45.2 GB at 4-bit | 45.2 GB | Too big |
| Llama 3.3 70B InstructMeta | needs 45.2 GB at 4-bit | 45.2 GB | Too big |
Memory needed = weights at that precision (real Hugging Face download sizes) + FP16 KV cache for 8,192 tokens. "Fits" means under 90% of 40 GB; "Tight" means 4-bit loads but leaves under 10% free. A * means the KV cache isn't sourced yet, so only the weights are counted. Data checked 2026-10-30.
Image generation, video editing and gaming
| Workload | Min–comfortable | Verdict |
|---|---|---|
| Stable Diffusion 1.5 (512×512)Image generation | 4–8 GB | Comfortable |
| SDXL 1.0 (1024×1024)Image generation | 8–12 GB | Comfortable |
| FLUX.1 [dev] (1024×1024)Image generation | 8–16 GB | Comfortable |
| Video editing, 1080p (HD)Video editing | 4–8 GB | Comfortable |
| Video editing, 4KVideo editing | 8–12 GB | Comfortable |
| Video editing, 8KVideo editing | 16–24 GB | Comfortable |
| Gaming, 1080pGaming | 8–12 GB | Comfortable |
| Gaming, 1440pGaming | 12–16 GB | Comfortable |
| Gaming, 4KGaming | 16 GB | Comfortable |
Ranges come from published guidance (Adobe, Puget Systems, AUTOMATIC1111 and others); the sources are linked in the calculator when you pick a workload.
Other memory sizes
8 GB12 GB16 GB24 GB32 GB40 GB48 GB64 GB96 GB128 GB
Want a different context length, more than one GPU, or your exact card? Use the What can my GPU run? tool. To check one model in detail, use the VRAM calculator.
Too big for your card? Rent the VRAM by the hour, or upgrade.
Run it on a cloud GPU
No upgrade needed
Rent a 24–80 GB GPU for the hour you need it.
Vast.ai 
Runpod 
E2E Networks Buy this GPU
Own the VRAM
Compare current GPU prices first on our GPU price table and measured speeds.
We may earn from links marked Affiliate partner. Disclosure.