32 GB · RTX 5090-class, 1792 GB/s · 32 GB system RAM

Qwen2.5 32B Instruct at Q4_K_M is the pick for 32 GB: it fits at 8K context, with about 32.8B active parameters, the most of any catalog model that clears the line.

Gaming PC with a large graphics card beside a monitor running a local AI chat

Qwen2.5 32B Instruct uses 22.0 GiB and runs at 45–72 tok/s. Active parameters are about 33B. Model page.

Sorted with Fits first. The pick is the Fits row with the most active parameters.
ModelQuantWeightsTotalVerdictSpeedOllama
Qwen3.6 35B-A3BUD-Q4_K_M20.6 GiB22.4 GiBFits414–668 tok/s—
Qwen2.5 32B InstructQ4_K_M18.5 GiB22.0 GiBFits45–72 tok/sollama run qwen2.5:32b
DeepSeek-R1 Distill Qwen 32BQ4_K_M18.5 GiB22.0 GiBFits45–72 tok/sollama run deepseek-r1:32b
Qwen3 32BQ4_K_M18.4 GiB22.0 GiBFits45–73 tok/sollama run qwen3:32b
Gemma 4 31BQ4_K_M17.1 GiB19.9 GiBFits50–78 tok/sollama run gemma4:31b
Qwen3 30B-A3BQ4_K_M17.3 GiB19.5 GiBFits347–706 tok/sollama run qwen3:30b-a3b
Qwen3.8 27BUD-Q4_K_M15.3 GiB17.3 GiBFits57–87 tok/s—
Gemma 3 27B InstructQ4_K_M15.4 GiB17.8 GiBFits56–87 tok/sollama run gemma3:27b
Gemma 4 26B-A4BUD-Q4_K_M15.8 GiB17.5 GiBFits374–635 tok/sollama run gemma4:26b-a4b
Mistral Small 3.2 24B InstructQ4_K_M13.3 GiB15.9 GiBFits63–100 tok/s—
Devstral Small 2 24B InstructQ4_K_M13.3 GiB15.9 GiBFits63–100 tok/s—
gpt-oss-20bMXFP411.3 GiB12.6 GiBFits375–592 tok/sollama run gpt-oss:20b
Qwen2.5 14B InstructQ4_K_M8.37 GiB10.9 GiBFits93–159 tok/sollama run qwen2.5:14b
DeepSeek-R1 Distill Qwen 14BQ4_K_M8.37 GiB10.9 GiBFits93–159 tok/sollama run deepseek-r1:14b
Qwen3 14BQ4_K_M8.38 GiB10.6 GiBFits95–159 tok/sollama run qwen3:14b
Phi-4Q4_K_M8.43 GiB11.0 GiBFits92–158 tok/sollama run phi4:14b
Mistral Nemo 12B InstructQ4_K_M6.96 GiB9.15 GiBFits112–192 tok/sollama run mistral-nemo:12b
Gemma 4 12BQ4_K_M6.63 GiB7.95 GiBFits130–201 tok/sollama run gemma4:12b
Gemma 3 12B InstructQ4_K_M6.80 GiB8.52 GiBFits121–196 tok/sollama run gemma3:12b
Qwen3.5 9BQ4_K_M5.29 GiB6.39 GiBFits164–252 tok/sollama run qwen3.5:9b
Qwen3 8BQ4_K_M4.68 GiB6.62 GiBFits158–285 tok/sollama run qwen3:8b
Llama 3.1 8B InstructQ4_K_M4.58 GiB6.39 GiBFits164–291 tok/sollama run llama3.1:8b
Qwen2.5 7B InstructQ4_K_M4.36 GiB5.56 GiBFits191–306 tok/sollama run qwen2.5:7b
Gemma 4 E4BQ4_K_M4.64 GiB5.66 GiBFits188–288 tok/sollama run gemma4:e4b
Gemma 4 E2BQ4_K_M2.89 GiB3.69 GiBFits304–461 tok/sollama run gemma4:e2b
Gemma 3 4B InstructQ4_K_M2.32 GiB3.24 GiBFits355–576 tok/sollama run gemma3:4b
Phi-4-mini InstructQ4_K_M2.32 GiB4.01 GiBFits276–575 tok/sollama run phi4-mini
Llama 3.2 3B InstructQ4_K_M1.88 GiB3.42 GiBFits333–710 tok/sollama run llama3.2:3b
SmolLM3 3BQ4_K_M1.78 GiB2.99 GiBFits391–748 tok/s—
Llama 3.2 1B InstructQ4_K_M0.75 GiB1.58 GiBFits916–1775 tok/sollama run llama3.2:1b
Qwen2.5 72B InstructQ4_K_M44.2 GiB49.5 GiBOffload2 tok/sollama run qwen2.5:72b
Llama 3.3 70B InstructQ4_K_M39.6 GiB44.7 GiBOffload2–3 tok/sollama run llama3.3:70b
gpt-oss-120bMXFP459.0 GiB62.8 GiBNo—ollama run gpt-oss:120b

FAQ

What does best mean?

The model that fits, with the most active parameters. Dense models count every parameter. MoE models count the experts that actually run, plus attention and embeddings. It is not a quality benchmark.

Which PC is this?

32 GB of VRAM at RTX 5090-class, 1792 GB/s, plus 32 GB of system RAM. A card with the same VRAM and less bandwidth fits the same models and runs them slower.

Why might a tighter model look smarter?

Tight still loads. It is kept out of the pick because an 8K chat can tip it into swapping. The table shows those models under the pick.