Can I run Qwen3.8 27B on the Apple M4?
Runs slowlyQ4_K_M runs entirely from 18.0 GB of system RAM: Runs slowly, est. 3.3 tok/s (2.3–4.3, theoretical estimate ±30%).
24 GB unified memory · 120.0 GB/s · FP16 3.8 TFLOPS. Assumes 8k context and an f16 KV cache. Verdicts are shown for each memory size.
Verdict by memory size
The page uses the 24 GB configuration; here is every size.
- 16 GBWon't runTry Q2_K (heavy quality loss): Runs slowlyest. 5.3 tok/s (3.7–6.9, theoretical estimate ±30%)
- 24 GBRuns slowlyest. 3.3 tok/s (2.3–4.3, theoretical estimate ±30%)(used on this page)
- 32 GBRuns slowlyest. 5.1 tok/s (4.1–6.1, calibrated estimate ±20%)
Every quant of Qwen3.8 27B on the Apple M4
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| Q2_K | 10.82 GB | Runs slowly | est. 8.0 tok/s6.4–9.6calibrated estimate ±20% | 11.9 / 16.8 GB | Unified memory | — |
| IQ4_XS | 15.48 GB | Runs slowly | est. 5.7 tok/s4.6–6.8calibrated estimate ±20% | 16.6 / 16.8 GB | Unified memory | — |
| Q4_K_Mbaseline | 17.44 GB | Runs slowly | est. 3.3 tok/s2.3–4.3theoretical estimate ±30% | 18.0 GB RAM | CPU only | — |
| Q8_0 | 29.12 GB | Won't run | — | Needs about 30.2 GB; 20.0 GB of memory is free (16.8 GB usable by the GPU) | — | — |
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Where the memory goes at Q4_K_M
- Weights in system RAM
- 18.0 GB
Context length vs. KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 17.7 GBRuns slowly est. 3.4 tok/s2.4–4.4theoretical estimate ±30% | 17.6 GBRuns slowly est. 3.4 tok/s2.4–4.4theoretical estimate ±30% | 17.5 GBRuns slowly est. 3.4 tok/s2.4–4.5theoretical estimate ±30% |
| 8k | 18.0 GBRuns slowly est. 3.3 tok/s2.3–4.3theoretical estimate ±30% | 17.7 GBRuns slowly est. 3.4 tok/s2.4–4.4theoretical estimate ±30% | 17.6 GBRuns slowly est. 3.4 tok/s2.4–4.4theoretical estimate ±30% |
| 16k | 18.5 GBRuns slowly est. 3.2 tok/s2.3–4.2theoretical estimate ±30% | 18.0 GBRuns slowly est. 3.3 tok/s2.3–4.3theoretical estimate ±30% | 17.7 GBRuns slowly est. 3.4 tok/s2.4–4.4theoretical estimate ±30% |
| 32k | 19.6 GBRuns slowly est. 3.1 tok/s2.1–4.0theoretical estimate ±30% | 18.6 GBRuns slowly est. 3.2 tok/s2.3–4.2theoretical estimate ±30% | 18.1 GBRuns slowly est. 3.3 tok/s2.3–4.3theoretical estimate ±30% |
| 64k | 22.8 GBdoes not fit | 19.7 GBRuns slowly est. 3.0 tok/s2.1–4.0theoretical estimate ±30% | 18.7 GBRuns slowly est. 3.2 tok/s2.3–4.2theoretical estimate ±30% |
| 128k | 27.7 GBdoes not fit | 23.7 GBdoes not fit | 19.9 GBRuns slowly est. 3.0 tok/s2.1–3.9theoretical estimate ±30% |
| 256k | 37.5 GBdoes not fit | 29.5 GBdoes not fit | 25.2 GBdoes not fit |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At Q4_K_M the verdict stays Runs slowly up to 32k context with an f16 KV cache, and up to 64k with q8_0.
Estimated speed
- Generation
- est. 3.3 tok/s (2.3–4.3, theoretical estimate ±30%)
- Prompt processing
- Prompt-processing estimates only apply when the whole model runs on the GPU.
Measured on this exact combination
No public measurement for this exact combination yet; the numbers above are estimates.
If this is not enough
A smaller model that runs well on the Apple M4
Runs greatgpt-oss-20b: Runs great, est. 31.9 tok/s (25.5–38.3, calibrated estimate ±20%)
Qwen3.8 27B on the Apple M5 Pro
Runs wellQ4_K_M: Runs well, est. 13.0 tok/s (10.4–15.6, calibrated estimate ±20%)
How to run it
Commands for the Q4_K_M file. Both tools download from Hugging Face on first run.
File: Qwen3.8-27B-Q4_K_M.gguf (17.44 GB) from bartowski/Qwen3.8-27B-GGUF.
llama-server -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M -c 8192 -ngl 0 -fa on- -ngl 0 keeps everything on the CPU.
- -fa on enables flash attention (needed for KV cache quantization).
- --cache-type-k q8_0 --cache-type-v q8_0 shrinks the KV cache to 0.29 GB at 8k context.
Terminal:
OLLAMA_CONTEXT_LENGTH=8192 ollama serve
ollama run hf.co/bartowski/Qwen3.8-27B-GGUF:Q4_K_M- Ollama defaults to 4k context on GPUs under 24 GB. Set OLLAMA_CONTEXT_LENGTH when starting the server, or /set parameter num_ctx inside the chat.
/set parameter num_ctx 8192
Related pages
Guide: Running local LLMs on a Mac (M4, M5): how much unified memory?
Other models on the Apple M4
Qwen3.8 27B on other hardware
Frequently asked questions
Can I run Qwen3.8 27B on the Apple M4?
Q4_K_M runs entirely from 18.0 GB of system RAM: Runs slowly, est. 3.3 tok/s (2.3–4.3, theoretical estimate ±30%). Q4_K_M needs 18.0 GB at 8k context; the GPU can use 16.8 GB of the unified memory.
How much context can Qwen3.8 27B use on the Apple M4?
At Q4_K_M the verdict stays Runs slowly up to 32k context with an f16 KV cache (2.1 GB of KV) and up to 64k with a q8_0 KV cache (2.3 GB). The model supports up to 256k.
Which quant should I use, and how fast is it?
Q4_K_M (17.44 GB) is the recommended balance: Runs slowly, est. 3.3 tok/s (2.3–4.3, theoretical estimate ±30%). It is also the largest tracked file that gets this verdict.
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.