8 GB · RTX 4060-class, 272 GB/s · 32 GB system RAM

Qwen3.5 9B at Q4_K_M is the pick for 8 GB: it fits at 8K context, with about 9B 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

Qwen3.5 9B uses 6.39 GiB and runs at 25–38 tok/s. Active parameters are about 9.0B. Model page.

Sorted with Fits first. The pick is the Fits row with the most active parameters.
ModelQuantWeightsTotalVerdictSpeedOllama
Qwen3.5 9BQ4_K_M5.29 GiB6.39 GiBFits25–38 tok/sollama run qwen3.5:9b
Qwen3 8BQ4_K_M4.68 GiB6.62 GiBFits24–43 tok/sollama run qwen3:8b
Llama 3.1 8B InstructQ4_K_M4.58 GiB6.39 GiBFits25–44 tok/sollama run llama3.1:8b
Qwen2.5 7B InstructQ4_K_M4.36 GiB5.56 GiBFits29–46 tok/sollama run qwen2.5:7b
Gemma 4 E4BQ4_K_M4.64 GiB5.66 GiBFits29–44 tok/sollama run gemma4:e4b
Gemma 4 E2BQ4_K_M2.89 GiB3.69 GiBFits46–70 tok/sollama run gemma4:e2b
Gemma 3 4B InstructQ4_K_M2.32 GiB3.24 GiBFits54–87 tok/sollama run gemma3:4b
Phi-4-mini InstructQ4_K_M2.32 GiB4.01 GiBFits42–87 tok/sollama run phi4-mini
Llama 3.2 3B InstructQ4_K_M1.88 GiB3.42 GiBFits51–108 tok/sollama run llama3.2:3b
SmolLM3 3BQ4_K_M1.78 GiB2.99 GiBFits59–114 tok/s—
Llama 3.2 1B InstructQ4_K_M0.75 GiB1.58 GiBFits139–269 tok/sollama run llama3.2:1b
Gemma 4 12BQ4_K_M6.63 GiB7.95 GiBTight20–31 tok/sollama run gemma4:12b
Qwen3.6 35B-A3BUD-Q4_K_M20.6 GiB22.4 GiBOffload9–13 tok/s—
Qwen2.5 32B InstructQ4_K_M18.5 GiB22.0 GiBOffload1 tok/sollama run qwen2.5:32b
DeepSeek-R1 Distill Qwen 32BQ4_K_M18.5 GiB22.0 GiBOffload1 tok/sollama run deepseek-r1:32b
Qwen3 32BQ4_K_M18.4 GiB22.0 GiBOffload1 tok/sollama run qwen3:32b
Gemma 4 31BQ4_K_M17.1 GiB19.9 GiBOffload1–2 tok/sollama run gemma4:31b
Qwen3 30B-A3BQ4_K_M17.3 GiB19.5 GiBOffload10–14 tok/sollama run qwen3:30b-a3b
Qwen3.8 27BUD-Q4_K_M15.3 GiB17.3 GiBOffload1–2 tok/s—
Gemma 3 27B InstructQ4_K_M15.4 GiB17.8 GiBOffload1–2 tok/sollama run gemma3:27b
Gemma 4 26B-A4BUD-Q4_K_M15.8 GiB17.5 GiBOffload10–14 tok/sollama run gemma4:26b-a4b
Mistral Small 3.2 24B InstructQ4_K_M13.3 GiB15.9 GiBOffload2 tok/s—
Devstral Small 2 24B InstructQ4_K_M13.3 GiB15.9 GiBOffload2 tok/s—
gpt-oss-20bMXFP411.3 GiB12.6 GiBOffload13–19 tok/sollama run gpt-oss:20b
Qwen2.5 14B InstructQ4_K_M8.37 GiB10.9 GiBOffload4–6 tok/sollama run qwen2.5:14b
DeepSeek-R1 Distill Qwen 14BQ4_K_M8.37 GiB10.9 GiBOffload4–6 tok/sollama run deepseek-r1:14b
Qwen3 14BQ4_K_M8.38 GiB10.6 GiBOffload4–6 tok/sollama run qwen3:14b
Phi-4Q4_K_M8.43 GiB11.0 GiBOffload4–6 tok/sollama run phi4:14b
Mistral Nemo 12B InstructQ4_K_M6.96 GiB9.15 GiBOffload8–12 tok/sollama run mistral-nemo:12b
Gemma 3 12B InstructQ4_K_M6.80 GiB8.52 GiBOffload15–21 tok/sollama run gemma3:12b
gpt-oss-120bMXFP459.0 GiB62.8 GiBNo—ollama run gpt-oss:120b
Qwen2.5 72B InstructQ4_K_M44.2 GiB49.5 GiBNo—ollama run qwen2.5:72b
Llama 3.3 70B InstructQ4_K_M39.6 GiB44.7 GiBNo—ollama run llama3.3:70b

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?

8 GB of VRAM at RTX 4060-class, 272 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.