Can I run Qwen3.5 9B on the Apple M4?
Runs wellQ4_K_M fits in unified memory (6.7 of 11.2 GB usable by the GPU): Runs well, est. 14.9 tok/s (11.9–17.8, calibrated estimate ±20%).
16 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 16 GB configuration; here is every size.
- 16 GBRuns wellest. 14.9 tok/s (11.9–17.8, calibrated estimate ±20%)(used on this page)
- 24 GBRuns wellest. 14.9 tok/s (11.9–17.8, calibrated estimate ±20%)
- 32 GBRuns wellest. 14.9 tok/s (11.9–17.8, calibrated estimate ±20%)
Every quant of Qwen3.5 9B on the Apple M4
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| IQ4_XS | 5.47 GB | Runs well | est. 15.9 tok/s12.7–19.1calibrated estimate ±20% | 6.3 / 11.2 GB | Unified memory | — |
| Q4_K_Mbaseline | 5.87 GB | Runs well | est. 14.9 tok/s11.9–17.8calibrated estimate ±20% | 6.7 / 11.2 GB | Unified memory | — |
| Q8_0 | 9.79 GB | Runs slowly | est. 9.1 tok/s7.3–10.9calibrated estimate ±20% | 10.6 / 11.2 GB | Unified memory | — |
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
- 5.9 GB
- KV cache
- 0.3 GB
- Compute buffer
- 0.6 GB
- Free
- 4.5 GB
- Not usable by the GPU
- 4.8 GB
Context length vs. KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 6.5 GBRuns well est. 15.2 tok/s12.2–18.2calibrated estimate ±20% | 6.5 GBRuns well est. 15.3 tok/s12.3–18.4calibrated estimate ±20% | 6.4 GBRuns well est. 15.4 tok/s12.3–18.5calibrated estimate ±20% |
| 8k | 6.7 GBRuns well est. 14.9 tok/s11.9–17.8calibrated estimate ±20% | 6.6 GBRuns well est. 15.2 tok/s12.1–18.2calibrated estimate ±20% | 6.5 GBRuns well est. 15.3 tok/s12.3–18.4calibrated estimate ±20% |
| 16k | 7.1 GBRuns well est. 14.2 tok/s11.4–17.1calibrated estimate ±20% | 6.8 GBRuns well est. 14.8 tok/s11.8–17.8calibrated estimate ±20% | 6.7 GBRuns well est. 15.1 tok/s12.1–18.2calibrated estimate ±20% |
| 32k | 7.7 GBRuns well est. 13.1 tok/s10.5–15.8calibrated estimate ±20% | 7.2 GBRuns well est. 14.2 tok/s11.3–17.0calibrated estimate ±20% | 7.0 GBRuns well est. 14.8 tok/s11.8–17.7calibrated estimate ±20% |
| 64k | 9.1 GBRuns slowly est. 11.4 tok/s9.1–13.7calibrated estimate ±20% | 8.1 GBRuns well est. 13.0 tok/s10.4–15.6calibrated estimate ±20% | 7.6 GBRuns well est. 14.1 tok/s11.3–16.9calibrated estimate ±20% |
| 128k | 10.2 GBRuns slowly est. 5.9 tok/s4.1–7.7theoretical estimate ±30% | 9.9 GBRuns slowly est. 11.2 tok/s8.9–13.4calibrated estimate ±20% | 8.8 GBRuns well est. 12.9 tok/s10.3–15.4calibrated estimate ±20% |
| 256k | 17.4 GBdoes not fit | 10.5 GBRuns slowly est. 5.7 tok/s4.0–7.5theoretical estimate ±30% | 8.3 GBRuns slowly est. 7.2 tok/s5.0–9.4theoretical estimate ±30% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At Q4_K_M the verdict stays Runs well up to 32k context with an f16 KV cache, and up to 64k with q8_0.
Estimated speed
- Generation
- est. 14.9 tok/s (11.9–17.8, calibrated estimate ±20%)
- Prompt processing
- about 152 tok/s (91–213, rough estimate ±40%)
Measured on this exact combination
No public measurement for this exact combination yet; the numbers above are estimates.
If this is not enough
A larger model that also runs well on the Apple M4
Runs wellGemma 3 12B: Runs well, est. 11.6 tok/s (9.3–14.0, calibrated estimate ±20%)
Qwen3.5 9B on the Apple M4 Pro
Runs greatQ4_K_M: Runs great, 37.0 tok/s measured at 4k context (1 run; 8k estimate 33.8 tok/s, 27.0–40.6, calibrated estimate ±20%)
How to run it
Commands for the Q4_K_M file. Both tools download from Hugging Face on first run.
File: Qwen3.5-9B-Q4_K_M.gguf (5.87 GB) from unsloth/Qwen3.5-9B-MTP-GGUF.
llama-server -hf unsloth/Qwen3.5-9B-MTP-GGUF:Q4_K_M -c 8192 -ngl all -fa on- -ngl all loads every layer on the GPU.
- -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 0.14 GB at 8k context.
Terminal:
OLLAMA_CONTEXT_LENGTH=8192 ollama serve
ollama run hf.co/unsloth/Qwen3.5-9B-MTP-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
Qwen3.5 9B on other hardware
Frequently asked questions
Can I run Qwen3.5 9B on the Apple M4?
Q4_K_M fits in unified memory (6.7 of 11.2 GB usable by the GPU): Runs well, est. 14.9 tok/s (11.9–17.8, calibrated estimate ±20%). Q4_K_M needs 6.7 GB at 8k context; the GPU can use 11.2 GB of the unified memory.
How much context can Qwen3.5 9B use on the Apple M4?
At Q4_K_M the verdict stays Runs well up to 32k context with an f16 KV cache (1.1 GB of KV) and up to 64k with a q8_0 KV cache (1.1 GB). The model supports up to 256k.
Which quant should I use, and how fast is it?
Q4_K_M (5.87 GB) is the recommended balance: Runs well, est. 14.9 tok/s (11.9–17.8, calibrated estimate ±20%). 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.