What fits in 48 GB of VRAM
Data updated . Last reviewed the same day, when this catalog snapshot was written.
Usable memory here is 45.6 GB, which is 95% of 48 GB. 38 of 49 catalog models fit at Q4, Q8, or FP16. No GPU in this catalog ships with exactly 48 GB. The list is the memory budget, not a product roundup. 2 of the fits do not fit the 32 GB tier.
Models
Sorted by the highest quant that fits, then by headroom. Tok/s is blank because this tier has no catalog GPU.
| Model | Best quant | Estimate | Headroom | Est. tok/s |
|---|---|---|---|---|
| Llama 3.2 1B Instruct | FP16 | 3.0 GB | 42.6 GB | — |
| Llama 3.2 3B Instruct | FP16 | 7.7 GB | 37.9 GB | — |
| Phi-3.5 mini | FP16 | 9.2 GB | 36.4 GB | — |
| Phi-4 mini | FP16 | 9.2 GB | 36.4 GB | — |
| Gemma 3 4B | FP16 | 10.3 GB | 35.3 GB | — |
| Mistral 7B Instruct v0.3 | FP16 | 17.4 GB | 28.2 GB | — |
| Qwen2.5 7B Instruct | FP16 | 18.3 GB | 27.3 GB | — |
| Gemma 4 E4B | FP16 | 19.2 GB | 26.4 GB | — |
| DeepSeek R1 Distill Llama 8B | FP16 | 19.3 GB | 26.3 GB | — |
| Llama 3.1 8B Instruct | FP16 | 19.3 GB | 26.3 GB | — |
| Qwen3 8B | FP16 | 19.7 GB | 25.9 GB | — |
| Ministral 3 8B | FP16 | 21.4 GB | 24.2 GB | — |
| Gemma 2 9B | FP16 | 22.2 GB | 23.4 GB | — |
| Gemma 4 12B | FP16 | 28.7 GB | 16.9 GB | — |
| Gemma 3 12B | FP16 | 29.2 GB | 16.4 GB | — |
| Mistral Nemo 12B | FP16 | 29.4 GB | 16.2 GB | — |
| Ministral 3 14B | FP16 | 33.5 GB | 12.1 GB | — |
| Phi-4 | FP16 | 35.2 GB | 10.4 GB | — |
| Qwen3 14B | FP16 | 35.4 GB | 10.2 GB | — |
| Qwen2.5 14B Instruct | FP16 | 35.4 GB | 10.2 GB | — |
| DeepSeek V4 Flash | Q8 | 4.5 GB | 41.1 GB | — |
| gpt-oss-20b | Q8 | 12.3 GB | 33.3 GB | — |
| Mistral Small 3.2 24B | Q8 | 26.4 GB | 19.2 GB | — |
| Gemma 4 26B-A4B | Q8 | 27.2 GB | 18.4 GB | — |
| Gemma 3 27B | Q8 | 28.7 GB | 16.9 GB | — |
| Qwen3.6 27B | Q8 | 29.1 GB | 16.5 GB | — |
| Qwen3.8 27B | Q8 | 29.6 GB | 16.0 GB | — |
| Gemma 2 27B | Q8 | 30.0 GB | 15.6 GB | — |
| GLM-4.7-Flash | Q8 | 32.3 GB | 13.3 GB | — |
| Qwen3 30B-A3B | Q8 | 33.3 GB | 12.3 GB | — |
| Gemma 4 31B | Q8 | 34.0 GB | 11.6 GB | — |
| Qwen3 32B | Q8 | 37.0 GB | 8.6 GB | — |
| DeepSeek R1 Distill Qwen 32B | Q8 | 37.0 GB | 8.6 GB | — |
| Qwen2.5 32B Instruct | Q8 | 37.0 GB | 8.6 GB | — |
| Qwen3.6 35B-A3B | Q8 | 37.1 GB | 8.5 GB | — |
| Mixtral 8x7B | Q4 | 27.5 GB | 18.1 GB | — |
| Llama 3.3 70B Instruct | Q4 | 42.5 GB | 3.1 GB | — |
| DeepSeek R1 Distill Llama 70B | Q4 | 45.2 GB | 0.4 GB | — |
Next misses at Q4: Qwen2.5 72B Instruct (46.7 GB), gpt-oss-120b (63.1 GB), Llama 4 Scout (65.4 GB), Mistral Small 4 (74.0 GB), Mixtral 8x22B (87.5 GB).
GPUs at 48 GB
No catalog GPU has exactly 48 GB of VRAM.
Questions
- How many catalog models fit in 48 GB of VRAM?
- 38 of 49 fit at Q4 or better inside 45.6 GB usable (95% of 48 GB).
- Which catalog GPUs have 48 GB of VRAM?
- None. No GeForce or Radeon row in this catalog is exactly 48 GB.
- Is the tokens/sec column a benchmark?
- No column is filled in, because no catalog GPU has 48 GB to borrow a bandwidth from.