Best local LLMs for 32 GB of VRAM
On one 32 GB GPU at Q4_K_M with 32K context, the largest dense model that fits is Gemma 4 31B (31.3B, 23.9 GB, about 40–57 tokens/s on one RTX 5090), the largest MoE is K2-Horizon MoVA 36B-A4B (37.4B total, 4.0B active, 30.3 GB, about 56–96 tokens/s), among models that take 16 GB or more, the longest context goes to a tie between Ornith 1.5 35B-A3B and Qwen3.6 35B-A3B (the full 256K in 28.3 GB), and the largest at Q8_0 is Gemma 4 26B-A4B (MoE, 25.8B total, 4.0B active, 29.6 GB).
Updated , the latest date the data of a model listed here was checked; 55 models in total. Ordered by memory and size, not quality. Total = weights + FP16 KV cache for one request + 0.5 GB + 10% overhead; GB = GiB. A model counts as fitting when it leaves at least 0.5 GB free, as on the Can I run it? pages. Speeds are bandwidth estimates for one request with the 32K context full, on one RTX 5090 (1792 GB/s), not measurements. Not sure how much VRAM you have? Detect your GPU.
gpt-oss-20b is listed at MXFP4, the format it is published in, in the Q4_K_M lists and not in the Q8_0 list: its GGUF quants keep the experts in MXFP4, so every type is about the same size.
Top picks for 32 GB
| Pick | Model | Parameters | Active | VRAM | Tokens/s (RTX 5090) |
|---|---|---|---|---|---|
| Largest dense (Q4_K_M) | Gemma 4 31B | 31.3B | dense | 23.9 GB | 40–57 |
| Largest MoE (Q4_K_M) | K2-Horizon MoVA 36B-A4B (MoE) | 37.4B | 4.0B | 30.3 GB | 56–96 |
| Longest context, 16 GB+ models (Q4_K_M) | Ornith 1.5 35B-A3B (MoE) | 36.0B | 3.0B | 256K (full), 28.3 GB, tied with Qwen3.6 35B-A3B | 163–305 |
| Largest at Q8_0 | Gemma 4 26B-A4B (MoE) | 25.8B | 4.0B | 29.6 GB | 89–158 |
What is the best local LLM for 32 GB VRAM?
By size, the largest model that fits one 32 GB GPU at Q4_K_M with 32K context is K2-Horizon MoVA 36B-A4B (MoE, 37.4B total, 4.0B active, 30.3 GB, 1.7 GB spare); 29 of the 55 models we track fit. This page ranks by memory and size, not by benchmark scores, which our data does not include.
Can 32 GB run a 70B model?
Not on one card: Llama 3.1 70B (70.6B) needs 55.2 GB at Q4_K_M with 32K context, or 2 × 32 GB cards, and still 41.3 GB at IQ3_XXS.
Every step is on the Llama 3.1 70B VRAM page.
Is Q8_0 or a bigger model better on 32 GB?
On one 32 GB GPU with 32K context, the largest model at Q8_0 is Gemma 4 26B-A4B (MoE, 25.8B total, 4.0B active, 29.6 GB, about 89–158 tokens/s) and at Q4_K_M it is K2-Horizon MoVA 36B-A4B (MoE, 37.4B total, 4.0B active, 30.3 GB, about 56–96 tokens/s), 1.5× the total parameters. Q8_0 stores about 8.5 bits per weight and Q4_K_M 4.84; our data covers memory and speed, not output quality, so this page does not score the trade-off.
Bits per weight for every GGUF type are in GGUF quantization explained.
Newest models that fit 32 GB
Sorted by the date each model's data was added or checked, newest first (Q4_K_M, 32K context). New models appear here as they are added.
| Model | Added / checked | Parameters | Active | VRAM (GB) | Tokens/s |
|---|---|---|---|---|---|
| Spark-X2.5 4B | 4.1B | dense | 4.40 | 186–300 | |
| Ling 3.0 Tiny 7.9B-A1.3B (MoE) | 7.9B | 1.3B | 5.62 | 295–613 | |
| LFM2.5 8B-A1B (MoE) | 8.5B | 1.5B | 6.16 | 254–510 | |
| Limite 1B Violetto | 1.0B | dense | 1.62 | 383–761 | |
| ZDTaichu 5.0 9B | 9.8B | dense | 7.67 | 116–175 | |
| Hemmingway-1 27B | 27.3B | dense | 19.63 | 49–69 | |
| MiMo V2.6 Distill Qwen 9B | 9.4B | dense | 7.43 | 120–181 | |
| K2-Horizon MoVA 36B-A4B (MoE) | 37.4B | 4.0B | 30.31 | 56–96 |
All 29 models that fit 32 GB at Q4_K_M
| Model | Parameters | Active | VRAM (GB) | Spare (GB) | Tokens/s | Calculator |
|---|---|---|---|---|---|---|
| K2-Horizon MoVA 36B-A4B (MoE) | 37.4B | 4.0B | 30.31 | 1.69 | 56–96 | Open |
| Ornith 1.5 35B-A3B (MoE) | 36.0B | 3.0B | 23.47 | 8.53 | 163–305 | Open |
| Qwen3.6 35B-A3B (MoE) | 36.0B | 3.0B | 23.47 | 8.53 | 163–305 | Open |
| LLM-jp-4.1 32B-A3B Thinking (MoE) | 32.1B | 3.8B | 22.62 | 9.38 | 102–182 | Open |
| Nemotron 3 Nano 30B-A3B (MoE) | 31.6B | 3.5B | 20.28 | 11.72 | 172–324 | Open |
| Gemma 4 31B | 31.3B | dense | 23.92 | 8.08 | 40–57 | Open |
| GLM-4.7 Flash | 31.2B | 3.0B | 21.67 | 10.33 | 122–222 | Open |
| Xing 4.0 29B-A4B (MoE) | 31.2B | 4.0B | 21.39 | 10.61 | 114–205 | Open |
| Qwen3 30B-A3B (MoE) | 30.5B | 3.3B | 22.72 | 9.28 | 89–158 | Open |
| Muse Glimmer 30B | 29.8B | dense | 19.51 | 12.49 | 49–70 | Open |
| Qwen3.6 27B | 27.8B | dense | 19.92 | 12.08 | 48–68 | Open |
| Qwen3.8 27B | 27.8B | dense | 19.92 | 12.08 | 48–68 | Open |
| Hemmingway-1 27B | 27.3B | dense | 19.63 | 12.37 | 49–69 | Open |
| Gemma 4 26B-A4B (MoE) | 25.8B | 4.0B | 17.50 | 14.50 | 128–233 | Open |
| gpt-oss-20b MXFP4 | 20.9B | 3.6B | 15.44 | 16.56 | 134–246 | Open |
| Gemma 4 12B | 12.0B | dense | 8.98 | 23.02 | 101–150 | Open |
| ZDTaichu 5.0 9B | 9.8B | dense | 7.67 | 24.33 | 116–175 | Open |
| Ornith 1.5 9B | 9.7B | dense | 7.58 | 24.42 | 117–177 | Open |
| Qwen3.5 9B | 9.7B | dense | 7.58 | 24.42 | 117–177 | Open |
| MiMo V2.6 Distill Qwen 9B | 9.4B | dense | 7.43 | 24.57 | 120–181 | Open |
| LFM2.5 8B-A1B (MoE) | 8.5B | 1.5B | 6.16 | 25.84 | 254–510 | Open |
| Qwen3 8B | 8.2B | dense | 10.53 | 21.47 | 87–128 | Open |
| Llama 3.1 8B | 8.0B | dense | 9.88 | 22.12 | 93–137 | Open |
| Gemma 4 E4B | 8.0B | dense | 6.05 | 25.95 | 143–221 | Open |
| Ling 3.0 Tiny 7.9B-A1.3B (MoE) | 7.9B | 1.3B | 5.62 | 26.38 | 295–613 | Open |
| Spark-X2.5 4B | 4.1B | dense | 4.40 | 27.60 | 186–300 | Open |
| Nemotron 3 Nano 4B | 4.0B | dense | 3.51 | 28.49 | 223–372 | Open |
| MiniCPM5 2B | 2.5B | dense | 3.50 | 28.50 | 223–373 | Open |
| Limite 1B Violetto | 1.0B | dense | 1.62 | 30.38 | 383–761 | Open |
All 15 models that fit 32 GB at Q8_0
| Model | Parameters | Active | VRAM (GB) | Spare (GB) | Tokens/s | Calculator |
|---|---|---|---|---|---|---|
| Gemma 4 26B-A4B (MoE) | 25.8B | 4.0B | 29.60 | 2.40 | 89–158 | Open |
| Gemma 4 12B | 12.0B | dense | 14.58 | 17.42 | 65–93 | Open |
| ZDTaichu 5.0 9B | 9.8B | dense | 12.26 | 19.74 | 76–111 | Open |
| Ornith 1.5 9B | 9.7B | dense | 12.11 | 19.89 | 77–112 | Open |
| Qwen3.5 9B | 9.7B | dense | 12.11 | 19.89 | 77–112 | Open |
| MiMo V2.6 Distill Qwen 9B | 9.4B | dense | 11.84 | 20.16 | 79–114 | Open |
| LFM2.5 8B-A1B (MoE) | 8.5B | 1.5B | 10.13 | 21.87 | 192–367 | Open |
| Qwen3 8B | 8.2B | dense | 14.37 | 17.63 | 66–95 | Open |
| Llama 3.1 8B | 8.0B | dense | 13.64 | 18.36 | 69–100 | Open |
| Gemma 4 E4B | 8.0B | dense | 9.80 | 22.20 | 93–138 | Open |
| Ling 3.0 Tiny 7.9B-A1.3B (MoE) | 7.9B | 1.3B | 9.32 | 22.68 | 223–436 | Open |
| Spark-X2.5 4B | 4.1B | dense | 6.33 | 25.67 | 137–211 | Open |
| Nemotron 3 Nano 4B | 4.0B | dense | 5.38 | 26.62 | 158–247 | Open |
| MiniCPM5 2B | 2.5B | dense | 4.68 | 27.32 | 177–283 | Open |
| Limite 1B Violetto | 1.0B | dense | 2.11 | 29.89 | 323–600 | Open |
Macs and CPU-only PCs with about 32 GB to use
These machines leave a model between 32 GB and the next size up: on a Mac, the unified memory macOS lets the GPU use by default (about two thirds up to 32 GB, 75% above; sudo sysctl iogpu.wired_limit_mb raises it), on a CPU-only PC the RAM minus about 6 GB for the system. The lists above apply to them too; speed follows each one's bandwidth, and a CPU is also much slower at reading long prompts.
| Machine | Usable memory | Bandwidth | Models at Q4_K_M (8K) | Biggest at 32K | Tokens/s |
|---|---|---|---|---|---|
| M4 Pro Mac (48 GB) | 36 GB | 273 GB/s | 29 | K2-Horizon MoVA 36B-A4B | 9.1–15 |
| M5 Pro Mac (48 GB) | 36 GB | 307 GB/s | 29 | K2-Horizon MoVA 36B-A4B | 10–17 |
| M4 Max Mac (48 GB) | 36 GB | 546 GB/s | 29 | K2-Horizon MoVA 36B-A4B | 18–30 |
| M5 Max Mac (48 GB) | 36 GB | 614 GB/s | 29 | K2-Horizon MoVA 36B-A4B | 20–34 |
Other VRAM sizes
All numbers, including 8K and 128K context, are in the open dataset, with summary figures on LLM VRAM statistics; any other setting can be worked out in the LLM VRAM Calculator.