Run Qwen 3.5 9B on M3 Ultra

Qwen 3.5 9B runs at 81 tok/s on a 96 GB M3 Ultra, a community run reported by github.com (llama.cpp, Q4_K_M, 75-token chat prompt, three greedy 256-token continuations). It needs about 7 GB of unified memory. It is Alibaba's 9B dense model, released under the Apache 2.0 licence.

Speed
community81
tok/s
First Token
—
seconds
RAM Needed
7
GB minimum
Engine
llama.cpp
of this figure
Want to run Qwen 3.5 9B faster — or step up to bigger models? Find your Mac →

The figure

This is a community run: a real run, reported by its source and linked below. The index records it as given and never adjusts it.

MetricValue
Tokens per secondcommunity81 tok/s
Provenancecommunity run · github.com
QuantizationQ4_K_M
Enginellama.cpp
Context75-token chat prompt, three greedy 256-token continuations
Figure dated2026-11-07
Minimum RAM7 GB
Parameters9B
Chip bandwidth819 GB/s

Every figure for Qwen 3.5 9B on the M3 Ultra

Each row as the index holds it, with its source. “Mac RAM” is the memory of the Mac a run used, not the model’s requirement (7 GB). The headline figure follows the reference rule: the lower-median measured plain-decode run at 5 bits or fewer (a run whose quant was not stated counts only when no stated-quant run qualifies), else the estimate; a run marked “not used to anchor estimates” is listed here but never becomes the headline.

SpeedQuantEngineMac RAMContextSourceNote
community114.1 tok/s4-bitMLX96 GB75-token chat prompt, three greedy 256-token continuationsgithub.comM3 Ultra, 60-core GPU (819 GB/s, the same as the 80-core part), 96 GB, macOS 26.4.1. mlx-lm on MLX (versions in the repo's M3 evidence folder), Kapitsa's MLX 4-bit weights. Run by the Kapitsa engine's developers with their own harness.
community81 tok/sQ4_K_Mllama.cpp96 GB75-token chat prompt, three greedy 256-token continuationsgithub.comM3 Ultra, 60-core GPU (819 GB/s, the same as the 80-core part), 96 GB, macOS 26.4.1. llama.cpp (build in the repo's M3 evidence folder), unsloth Q4_K_M. Run by the Kapitsa engine's developers with their own harness.
Provenance key
sourced
Published figure — vendor or press review
estimated
community
Linked individual run — blog, forum or submission

Setup: run Qwen 3.5 9B on a Mac

Ollama — install it (brew install ollama or the app from ollama.com), then pull the model by its tag, checked on ollama.com on 2026-11-14:

ollama run qwen3.5:9b

MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:

pip install -U mlx-lm
mlx_lm.generate --model <repo-id> --prompt "Hello"

Downloads: Ollama page → · Hugging Face model card →

The 81 tok/s figure above came from llama.cpp at Q4_K_M; other runtimes and quants run at different speeds.

Prefer a GUI? LM Studio downloads GGUF and MLX builds in one click. Step by step: install Ollama on a Mac.

Qwen 3.5 9B on other Apple Silicon chips

ChipSpeedQuantEngineBandwidth
M5 Ultracommunity167 tok/s4-bitMLX1228 GB/s
M5 Maxestimated81 tok/sQ4_K_MOllama614 GB/s
M4 Procommunity39 tok/sQ4_K_Mllama.cpp273 GB/s
M6estimated30 tok/sQ4_K_MMLX170 GB/s
M4estimated18 tok/sQ4_K_MLM Studio120 GB/s
M3estimated15 tok/sQ4_K_MOllama100 GB/s
M1estimated10 tok/sQ4_K_MOllama68 GB/s

For the estimated figures here, the bandwidth formula’s typical error for dense models is ±9%, against 75 measured runs of 20 models (how it is computed).

About Qwen 3.5 9B

Alibaba's 9B model, ~81 tok/s on M5 Max. Apache 2.0 license, LLMCheck capability 18/50.

Apache 2.0 licence · 9B · 262K context · LLMCheck capability 18/50 · see its leaderboard rank →

Memory by context length

Estimated, not measured: the index’s minimum for Qwen 3.5 9B (7 GB — the weights at the quant that fits, plus ~2 GB) plus the KV cache for that many tokens, computed from its config.json. hybrid: full attention on 8 of 32 layers; Gated DeltaNet layers keep a constant state. A 16-bit KV cache is the default in llama.cpp, Ollama, LM Studio and mlx-lm; a q8 cache halves the KV part. Coding agents hold 64K tokens or more. Runtimes also add a buffer that grows with context, so leave a few GB spare. macOS lets the GPU use only part of a Mac’s memory by default (figures for an M3 Ultra Mac on macOS 15, the lower default; macOS 26 allows more): “fits” means within the GPU memory a Mac of that tier can use with the limit raised, a tier tagged raise GPU limit fits only after sudo sysctl iogpu.wired_limit_mb raises it (the tooltip has the exact command), and the last column is the smallest Mac memory size whose default limit holds it (the GPU memory limit).

ContextKV cache (16-bit)RAM neededRAM, q8 KVSmallest tier that fitsWithin the default GPU limit
8K~0.3 GB~7.1 GB~7.1 GB16 GB Macsame
32K~1.1 GB~7.9 GB~7.5 GB16 GB Macsame
64K (coding-agent context)~2.1 GB~9.0 GB~8.0 GB16 GB Macsame
128K~4.3 GB~11.2 GB~9.1 GB16 GB Mac raise GPU limit (macOS 15)18 GB Mac

System requirements

Macs with the M3 Ultra

🛒 Get a Mac that runs Qwen 3.5 9B

Qwen 3.5 9B needs about 7 GB of unified memory. These Apple Silicon Macs have enough to run it.

As an Amazon Associate, LLMCheck earns from qualifying purchases. These affiliate links cost you nothing extra and help keep the LLMCheck index free.

Compare More Models

See how Qwen 3.5 9B stacks up against other models on your specific Mac hardware.

Open Compare Tool Full Leaderboard