Apple silicon · 24 GB unified · 273 GB/s

Yes. Apple M4 Pro, 24 GB fits gpt-oss-20b at MXFP4, the largest catalog model that clears the fit line at 8K context.

Compact mini PC and a laptop on a desk, both running a local AI chat

Usable memory is 18 GB after a 6.0 GB OS reserve.Apple lists 273 GB/s for M4 Pro and up to 64 GB of unified memory. Source.

Recommended quant, 8,192 tokens, unified memory.
ModelQuantWeightsTotalVerdictSpeedOllama
gpt-oss-20bMXFP411.3 GiB12.6 GiBFits57–90 tok/sollama run gpt-oss:20b
Qwen2.5 14B InstructQ4_K_M8.37 GiB10.9 GiBFits14–24 tok/sollama run qwen2.5:14b
DeepSeek-R1 Distill Qwen 14BQ4_K_M8.37 GiB10.9 GiBFits14–24 tok/sollama run deepseek-r1:14b
Qwen3 14BQ4_K_M8.38 GiB10.6 GiBFits15–24 tok/sollama run qwen3:14b
Phi-4Q4_K_M8.43 GiB11.0 GiBFits14–24 tok/sollama run phi4:14b
Mistral Nemo 12B InstructQ4_K_M6.96 GiB9.15 GiBFits17–29 tok/sollama run mistral-nemo:12b
Gemma 4 12BQ4_K_M6.63 GiB7.95 GiBFits20–31 tok/sollama run gemma4:12b
Gemma 3 12B InstructQ4_K_M6.80 GiB8.52 GiBFits18–30 tok/sollama run gemma3:12b
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 GiBFits29–47 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 GiBFits54–88 tok/sollama run gemma3:4b
Phi-4-mini InstructQ4_K_M2.32 GiB4.01 GiBFits42–88 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 GiBFits60–114 tok/s—
Llama 3.2 1B InstructQ4_K_M0.75 GiB1.58 GiBFits140–270 tok/sollama run llama3.2:1b
Qwen3.8 27BUD-Q4_K_M15.3 GiB17.3 GiBTight9–13 tok/s—
Gemma 3 27B InstructQ4_K_M15.4 GiB17.8 GiBTight9–13 tok/sollama run gemma3:27b
Gemma 4 26B-A4BUD-Q4_K_M15.8 GiB17.5 GiBTight57–97 tok/sollama run gemma4:26b-a4b
Mistral Small 3.2 24B InstructQ4_K_M13.3 GiB15.9 GiBTight10–15 tok/s—
Devstral Small 2 24B InstructQ4_K_M13.3 GiB15.9 GiBTight10–15 tok/s—
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
Qwen3.6 35B-A3BUD-Q4_K_M20.6 GiB22.4 GiBNo——
Qwen2.5 32B InstructQ4_K_M18.5 GiB22.0 GiBNo—ollama run qwen2.5:32b
DeepSeek-R1 Distill Qwen 32BQ4_K_M18.5 GiB22.0 GiBNo—ollama run deepseek-r1:32b
Qwen3 32BQ4_K_M18.4 GiB22.0 GiBNo—ollama run qwen3:32b
Gemma 4 31BQ4_K_M17.1 GiB19.9 GiBNo—ollama run gemma4:31b
Qwen3 30B-A3BQ4_K_M17.3 GiB19.5 GiBNo—ollama run qwen3:30b-a3b

FAQ

What does the verdict for Apple M4 Pro, 24 GB mean?

Memory is the GGUF file, plus KV cache at 8,192 tokens, plus a 0.5 GiB engine allowance, then 5% headroom. Fits means that total is at most 85% of fast memory. Tight means it still fits. Offload means it needs system RAM. No means it does not fit in VRAM plus system RAM after an 8 GB Windows reserve. The formula is on the methodology page.

What memory did you count as usable?

Fast memory is the 24 GB pool minus a 6.0 GB OS reserve (25% of unified memory, at least 3 GB and at most 12 GB).

Why is speed a range?

Decode speed is memory bandwidth divided by the bytes read per token. The high end is 80% of that roofline on the weights alone. The low end is 55% once the KV cache at the working context is included. It is not a benchmark.

How should I run Qwen3.8 27B?

The table uses each model's recommended quant, usually Q4_K_M or the vendor's published quant. Qwen3.8 27B has no verified Ollama tag, so no command is shown.