NeuroGenesysWhat LLMs can I run on a 32 GB GPU?
Last updated • By Will Schulz
On a 32 GB GPU you can run 17 of the 24 open-weight models we track with an 8K context. The largest is Llama 3.1 70B at 2-bit (Q2_K), which needs 32.5 GB, and 6 run unquantized at 16-bit.
Models that fit a 32 GB GPU
| Model | Best fit | Memory | Longest context | |
|---|---|---|---|---|
| Llama 3.1 70B | 2-bit (Q2_K) | 32.5 GB | 8K | calculator |
| Llama 3.3 70B | 2-bit (Q2_K) | 32.5 GB | 8K | calculator |
| Mixtral 8x7B | Q4_K_M | 31.5 GB | 16K | calculator |
| Qwen3 32B | Q6_K | 30.9 GB | 8K | calculator |
| Qwen2.5 32B | Q6_K | 30.6 GB | 8K | calculator |
| Qwen3 30B-A3B | Q6_K | 27.6 GB | 32K | calculator |
| Gemma 3 27B | INT8 / Q8_0 | 32.1 GB | 8K | calculator |
| Gemma 2 27B | Q6_K | 26.3 GB | 32K | calculator |
| gpt-oss-20b | INT8 / Q8_0 | 24.2 GB | 128K | calculator |
| Qwen2.5 14B | INT8 / Q8_0 | 18.7 GB | 64K | calculator |
| Phi-4 (14B) | INT8 / Q8_0 | 18.8 GB | 64K | calculator |
| Gemma 3 12B | BF16 / FP16 | 27.0 GB | 64K | calculator |
| Gemma 2 9B | BF16 / FP16 | 22.1 GB | 64K | calculator |
| Qwen3 8B | BF16 / FP16 | 19.1 GB | 64K | calculator |
| Llama 3.1 8B | BF16 / FP16 | 18.6 GB | 64K | calculator |
| Qwen2.5 7B | BF16 / FP16 | 17.2 GB | 128K | calculator |
| Mistral 7B v0.3 | BF16 / FP16 | 17.0 GB | 64K | calculator |
Memory is weights + FP16 KV cache + runtime overhead for one request, as in the VRAM calculator, assuming the runtime allocating 95% of the card. Longest context is the largest of 4K–128K that fits at that quantization; it is only usable up to the model's own context window.
Models too large for a 32 GB GPU
Even at 2-bit (Q2_K) these need more memory: Llama 3.1 405B, Llama 4 Scout, Llama 4 Maverick, Qwen2.5 72B, Qwen3 235B-A22B, gpt-oss-120b, DeepSeek-V3 / R1.
Questions
What is the largest LLM that runs on a 32 GB GPU?
Of the models we track, Llama 3.1 70B (70.6B parameters) at 2-bit (Q2_K), which needs 32.5 GB with an 8K context. Larger models need more memory or several cards.
Can a 32 GB GPU run a model at full 16-bit precision?
Yes, 6 models: Gemma 3 12B, Gemma 2 9B, Qwen3 8B, Llama 3.1 8B, Qwen2.5 7B, Mistral 7B v0.3. Larger models need quantizing, usually to Q4_K_M (about 4.9 bits per weight), which loses little quality.
How much memory does a 32 GB GPU leave for context?
The KV cache grows with every token of context, so the longest context that fits depends on the model and quantization; the table lists it for each model, trying 4K, 8K, 16K, 32K, 64K and 128K tokens. These figures assume the runtime allocating 95% of the card.
