Qwen3 235B-A22B Thinking 2507 on Apple M2 Ultra (192 GB)

Can the Apple M2 Ultra (192 GB) run Qwen3 235B-A22B Thinking 2507?

Yes, it runs well (tight fit)

Runs well (tight fit). At Q3_K_M it needs about 129 GB of 144 GB usable, with an estimated speed of 42.8 tok/s.

Best quant: Q3_K_M
Generation speed: 42.8 tok/s
Prompt processing: 767 tok/s
First token (1K prompt): 1.3 s
Largest safe context: 84K

Best app to run it on the Apple M2 Ultra (192 GB)

On macOS

LM Studio

Desktop app · runs on llama.cpp, MLX · use the MLX build

Runs the MLX build, typically up to 10% faster than GGUF for a model with 22B active parameters on Apple silicon.
Point-and-click: search, download and chat in one window.

Open Qwen3 235B-A22B Thinking 2507 in LM Studio

Ollama One command with Ollama’s own build of this model.
llama.cpp Usually the first to support new model architectures.

Want an agent? Qwen3 235B-A22B Thinking 2507 supports tool use, so you can pair a llama.cpp or MLX LM server with one of these. Apps that run on this computer come first.

Coding agents

Personal assistants

How we choose: the files published for this model, your operating system and GPU, whether it fits fully in memory, and published speed tests (MLX vs llama.cpp on Apple silicon, vLLM vs llama.cpp vs Ollama on NVIDIA). Details.

Download Qwen3 235B-A22B Thinking 2507

Recommended for the Apple M2 Ultra (192 GB): Q3_K_M. Every other quantization has its own download link in the table below.

Quickest · Ollama

One command downloads a ready-to-run version and starts a chat.

Get Ollama at ollama.com/download (opens in a new tab) · model page (opens in a new tab)

ollama run qwen3:235b-thinking

Quantized files · GGUF

Works in LM Studio, llama.cpp, Jan and Ollama. No sign-in needed.

Download the Q3_K_M files (opens in a new tab)

huggingface.co/unsloth/Qwen3-235B-A22B-Thinking-2507-GGUF (opens in a new tab)

ollama run hf.co/unsloth/Qwen3-235B-A22B-Thinking-2507-GGUF

Original weights · Hugging Face

Full-precision files from Alibaba, for fine-tuning, vLLM, Transformers.

huggingface.co/Qwen/Qwen3-235B-A22B-Thinking-2507 (opens in a new tab)

No sign-in needed.

Official repo · GitHub

Reference code, inference scripts and docs from Alibaba. The weights themselves download from Hugging Face.

github.com/QwenLM/Qwen3 (opens in a new tab)

Qwen’s main repo, renamed from Qwen2.5 to Qwen3; it still documents Qwen2.5.

git clone https://github.com/QwenLM/Qwen3

Links checked 3 days ago against Hugging Face, Ollama and GitHub.

How it’s rated

Quality 76Great quality (A) · 76 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 88/100 (quality, fit, speed and memory use) sets the ranking.: How capable the model itself is at Q3_K_M, from its editorial quality tier scaled by the quantization or precision. It’s the same on any hardware that runs this variant. Scale: A+ 85+ · A 70+ · B 55+ · C 40+ · D under 40.
Runs Runs welltight fit: How well it runs on the Apple M2 Ultra (192 GB): speed, capped by how well it fits in memory. Runs great: 30+ tokens/s (or near the image, video or decision speed target) · Runs well: 10+ tokens/s · Slow: 3+ tokens/s · Very slow: under 3 · Won’t run: doesn’t fit even with offload. A tight fit or RAM offload caps it at Runs well; heavy offload caps it at Slow.

Match score

To order lists, each result also gets a match score out of 100 from four parts. It favors the most capable model that still runs well.

Model qualityEditorial quality tier of the model, scaled by the quantization or precision used.	34.4 / 45
Memory fitFull marks when the model fits in the share of unified memory the GPU can use; fewer when it goes past that limit.	24.0 / 25
SpeedGeneration speed against a 40 tokens/s target for LLMs, or render time against a target for image and video.	15.0 / 15
Memory useRewards using the memory you have, so a bigger, better variant beats a tiny one that leaves most memory idle.	15.0 / 15
Match score	88.4 / 100

The match score orders results; it is not shown as a grade. Quality and Runs are the two ratings in the tables.

Every quantization on the Apple M2 Ultra (192 GB)

Quant	Memory needed	Available	Speed	Runs	Quality	Download
`Q2_K`	108 GB	144 GB	52.7 tok/s	Runs great	63Good quality (B) · 63 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 83/100 (quality, fit, speed and memory use) sets the ranking.	GGUF · folder (opens in a new tab)
`Q3_K_M`	129 GB	144 GB	42.8 tok/s	Runs welltight fit	76Great quality (A) · 76 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 88/100 (quality, fit, speed and memory use) sets the ranking.	GGUF · folder (opens in a new tab)
`Q4_K_M`	157 GB	144 GB	28.9 tok/s	Runs wellover GPU memory limit	86Excellent quality (A+) · 86 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 76/100 (quality, fit, speed and memory use) sets the ranking.	GGUF · folder (opens in a new tab)
`Q5_K_M`	181 GB	144 GB	—	Won't run	88Excellent quality (A+) · 88 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 11/100 (quality, fit, speed and memory use) sets the ranking.	GGUF · folder (opens in a new tab)
`Q6_K`	206 GB	144 GB	—	Won't run	90Excellent quality (A+) · 90 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 12/100 (quality, fit, speed and memory use) sets the ranking.	GGUF · folder (opens in a new tab)
`Q8_0`	263 GB	144 GB	—	Won't run	91Excellent quality (A+) · 91 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 12/100 (quality, fit, speed and memory use) sets the ranking.	GGUF · folder (opens in a new tab)
`F16`	484 GB	144 GB	—	Won't run	91Excellent quality (A+) · 91 of 100: how capable the model is at F16, whatever it runs on. Match score 12/100 (quality, fit, speed and memory use) sets the ranking.	GGUF · folder (opens in a new tab) Original weights (opens in a new tab)

Memory figures assume 8K of context. Longer conversations or documents grow the KV cache.

Other hardware that runs it

86Excellent quality (A+) · 86 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 92/100 (quality, fit, speed and memory use) sets the ranking. Apple M3 Ultra (256 GB) Q4_K_M · 34.7 tok/s · Runs great
86Excellent quality (A+) · 86 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 88/100 (quality, fit, speed and memory use) sets the ranking. Apple M3 Ultra (512 GB) Q4_K_M · 34.7 tok/s · Runs great
63Good quality (B) · 63 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 79/100 (quality, fit, speed and memory use) sets the ranking. NVIDIA GB10 (128 GB) Q2_K · 23.2 tok/s · Runs well (tight fit)
63Good quality (B) · 63 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 68/100 (quality, fit, speed and memory use) sets the ranking. NVIDIA RTX PRO 6000 Blackwell Q2_K · 56.8 tok/s · Runs well (offloads to RAM)
63Good quality (B) · 63 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking. Apple M4 Max (128 GB) Q2_K · 32.3 tok/s · Runs well (over GPU memory limit)
63Good quality (B) · 63 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 64/100 (quality, fit, speed and memory use) sets the ranking. NVIDIA RTX Spark N1X (6144-core GPU) (128 GB) Q2_K · 21.5 tok/s · Runs well (over GPU memory limit)

Quick answers

How much memory does Qwen3 235B-A22B Thinking 2507 need on the M2 Ultra 192GB?

At Q3_K_M it needs about 129 GB. The M2 Ultra 192GB has 144 GB usable.

How fast is Qwen3 235B-A22B Thinking 2507 on the M2 Ultra 192GB?

About 42.8 tok/s at Q3_K_M, an estimate from the hardware’s memory bandwidth and the size of the model.

Which quantization of Qwen3 235B-A22B Thinking 2507 should I use on the M2 Ultra 192GB?

Q3_K_M. Of the versions that fit, it is the one that runs best; when two run equally well, the check prefers the higher-rated one.

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.