Can I run Qwen 3.5 397B-A17B on a Mac?

Yes: Qwen 3.5 397B-A17B runs on 4 of 54 Macs in the LLMCheck index. It needs about 210 GB at a 4K context, 212 GB at 64K. The smallest and cheapest current Mac that runs it at 15+ tok/s is the Mac Studio M5 Ultra 256 GB ($9,499; 62 tok/s community-measured). Best measured: 61.6 tok/s on an M5 Ultra (community).

RAM at 4K
210
GB
RAM at 64K
212
GB, 16-bit KV
Macs that run it
4
of 54 configs
Best measured
community61.6
tok/s · M5 Ultra

Qwen 3.5 397B-A17B on every Mac

One row per Mac configuration, grouped by product line. The build is the one each speed is for; the context column is the longest prompt that fits inside macOS’s default GPU memory limit with a 16-bit KV cache (“raised” where the limit has to be raised first; “full” when the whole window fits). Speeds are decode tok/s from the LLMCheck index, each with its provenance mark. For large mixture-of-experts models (over 40B total), the formula’s typical error is about 34% (the median gap to 70 measured plain-decode runs; it tends to undershoot (median −33%), and an estimate scaled from the model’s own measured run is usually closer; method).

Runs fits macOS’s default GPU limit, 15+ tok/sRuns, slowly fits, under 15 tok/sRaise GPU limit fits once the limit is raised (command below)Doesn't fit needs more memory than the Mac has
MacVerdictBuildSpeedMax contextModel page
MacBook Air
MacBook Air M5 16 GBDoesn't fit————
MacBook Air M5 24 GBDoesn't fit————
MacBook Air M5 32 GBDoesn't fit————
MacBook Air M1 8 GB refurbDoesn't fit————
MacBook Air M4 16 GB usedDoesn't fit————
MacBook Air M2 16 GB refurbDoesn't fit————
MacBook Air M3 16 GB refurbDoesn't fit————
MacBook Air M1 16 GB refurbDoesn't fit————
MacBook Air M4 24 GB usedDoesn't fit————
MacBook Air M2 24 GB refurbDoesn't fit————
MacBook Air M3 24 GB refurbDoesn't fit————
MacBook Pro
MacBook Pro M6 16 GBDoesn't fit————
MacBook Pro M5 Pro 24 GBDoesn't fit————
MacBook Pro M6 24 GBDoesn't fit————
MacBook Pro M6 32 GBDoesn't fit————
MacBook Pro M5 Pro 48 GBDoesn't fit————
MacBook Pro M5 Max 64 GBDoesn't fit————
MacBook Pro M5 Pro 64 GBDoesn't fit————
MacBook Pro M5 Max 128 GBDoesn't fit————
MacBook Pro M3 Max 64 GB refurbDoesn't fit————
MacBook Pro M3 Max 128 GB refurbDoesn't fit————
Mac mini
Mac mini M6 16 GBDoesn't fit————
Mac mini M5 Pro 24 GBDoesn't fit————
Mac mini M6 24 GBDoesn't fit————
Mac mini M6 32 GBDoesn't fit————
Mac mini M5 Pro 48 GBDoesn't fit————
Mac mini M5 Pro 64 GBDoesn't fit————
Mac mini M2 Pro 16 GB refurbDoesn't fit————
Mac mini M4 16 GB usedDoesn't fit————
Mac mini M4 Pro 24 GB usedDoesn't fit————
Mac mini M4 24 GB usedDoesn't fit————
Mac mini M2 Pro 32 GB refurbDoesn't fit————
Mac mini M4 Pro 48 GB usedDoesn't fit————
Mac mini M4 Pro 64 GB usedDoesn't fit————
iMac
iMac M6 16 GBDoesn't fit————
iMac M6 24 GBDoesn't fit————
iMac M6 32 GBDoesn't fit————
Mac Studio
Mac Studio M5 Max 36 GBDoesn't fit————
Mac Studio M5 Max 48 GBDoesn't fit————
Mac Studio M5 Max 64 GBDoesn't fit————
Mac Studio M5 Ultra 96 GBDoesn't fit————
Mac Studio M5 Max 128 GBDoesn't fit————
Mac Studio M5 Ultra 256 GBRuns4-bitcommunity62 tok/s256K (full)—
Mac Studio M5 Ultra 512 GBRuns4-bitcommunity62 tok/s256K (full)—
Mac Studio M2 Max 32 GB refurbDoesn't fit————
Mac Studio M1 Ultra 64 GB refurbDoesn't fit————
Mac Studio M4 Max 64 GB usedDoesn't fit————
Mac Studio M2 Max 64 GB refurbDoesn't fit————
Mac Studio M1 Ultra 128 GB refurbDoesn't fit————
Mac Studio M2 Ultra 128 GB refurbDoesn't fit————
Mac Studio M4 Max 128 GB usedDoesn't fit————
Mac Studio M2 Ultra 192 GB refurbDoesn't fit————
Mac Studio M3 Ultra 256 GB usedRaise GPU limit (macOS 15) command ↓Q4_K_Mestimated43 tok/s256K (full) raised—
Mac Studio M3 Ultra 512 GB refurbRunsQ4_K_Mestimated43 tok/s256K (full)—
Provenance key
sourced
Published figure — vendor or press review
estimated
community
Linked individual run — blog, forum or submission

Context memory is computed from its config.json (hybrid: full attention on 15 of 60 layers; Gated DeltaNet layers keep a constant state): about 2.0 GB of 16-bit KV cache at 64K tokens. A q8 cache (llama.cpp -ctk q8_0 -ctv q8_0, Ollama OLLAMA_KV_CACHE_TYPE=q8_0) halves that part. Runtimes also grow a compute buffer with context, so a fit within about 1 GB of the limit is optimistic.

Raising the GPU memory limit

macOS lets the GPU wire only part of unified memory by default (Metal’s recommended working set; M5 and M6 Macs run macOS 26 or later, whose default is higher), and Qwen 3.5 397B-A17B needs about 210 GB. On the Macs marked Raise GPU limit it fits in RAM but not inside that default. The default is a performance threshold, not a hard wall: past it, the runtime moves part of it to the CPU or shortens the context (llama.cpp, Ollama), runs it slowly (MLX), or stops with an out-of-memory error. One command raises the limit until the next restart (sudo sysctl iogpu.wired_limit_mb=0 restores the default sooner); it leaves macOS a reserve, and other apps get less memory while the model is loaded.

Qwen 3.5 397B-A17B: memory by quantization

Weights are 397B parameters at each format’s effective size (the index’s GB-per-billion table), plus about 2 GB for the runtime and a short context; longer contexts add the KV cache computed from the model’s config.json at 16 bits. It is a mixture-of-experts model: all 397B parameters sit in memory although only 17B are read per token. The first row is the catalog’s own figure (about 210 GB at a 4K context), the one every verdict on this page uses. “Smallest Mac” is the smallest memory size among the Macs Apple sells new that holds it at a 4K context inside that Mac’s default GPU limit. Estimates, not measurements.

BuildWeightsRAM at 4Kat 32Kat 64KSmallest Mac
As catalogued—210 GB211 GB212 GB256 GB
Q2_K135.0 GB137 GB138 GB139 GB256 GB
Q3_K_M194.5 GB197 GB197 GB198 GB256 GB
Q4_K_M226.3 GB228 GB229 GB230 GB512 GB (256 GB with a raised limit)
Q5_K_M281.9 GB284 GB285 GB286 GB512 GB
Q6_K325.5 GB328 GB328 GB329 GB512 GB
Q8_0420.8 GB423 GB424 GB425 GB512 GB
F16794.0 GB796 GB797 GB798 GBnone (over 512 GB)

Measured runs of Qwen 3.5 397B-A17B

Every published (filled square) and community (hollow circle) run the index holds for Qwen 3.5 397B-A17B: 0 published, 1 community, each linked to its source. LLMCheck runs no tests of its own. One row is one measurement session; on each chip the table above shows the lower median of the eligible runs (plain decode at 5 bits or fewer), and a binned or “not an anchor” run is shown but never used as a reference.

ChipRuntimeQuantDecodeContextDateSource
M5 Ultra 256 GBMLX4-bitcommunity61.6 tok/s512-token random prompt, 128 generated, mean of 32026-11-12bohemiansilicon.com

Recommended setup

The index has no verified Ollama tag for Qwen 3.5 397B-A17B. Start from its listing and copy the exact name of a build that fits: Search the Ollama library → · model card →

GGUF builds run in LM Studio or llama.cpp; MLX builds run in LM Studio or mlx-lm.

On the Mac Studio M5 Ultra (256 GB), the smallest current Mac that runs it inside the default GPU limit, the index figure is 62 tok/s (community run) at 4-bit; keep the context at or under 256K there, or use a q8 KV cache.

Step-by-step: the Ollama install guide.

🛒 Which Mac to get for Qwen 3.5 397B-A17B

Mac Studio M5 Ultra (256 GB) — smallest current Mac that runs it and cheapest at 15+ tok/s, $9,499 new (Apple list price, checked 2026-11-14), 62 tok/s (community run). Amazon does not stock the Mac Studio M5 Ultra (256 GB): Apple builds it to order.

FAQ: Qwen 3.5 397B-A17B on a Mac

Can I run Qwen 3.5 397B-A17B on a Mac?

Yes. Qwen 3.5 397B-A17B needs about 210 GB of unified memory at a short (4K-token) context, so it runs on 4 of the 54 Mac configurations in the LLMCheck index; 3 of them hold it inside macOS’s default GPU memory limit and 1 only after raising that limit. The smallest Mac Apple sells new that runs it is the Mac Studio M5 Ultra (256 GB).

How much RAM does Qwen 3.5 397B-A17B need?

About 210 GB at a 4K context: the weights plus about 2 GB for the runtime. Context adds KV cache: about 211 GB in total at 32K and 212 GB at 64K with the default 16-bit cache (about 211 GB at 64K with a q8 cache). These are estimates from the model’s config.json, not measurements.

How fast does Qwen 3.5 397B-A17B run on a Mac?

On an M5 Ultra, the chip the leaderboard quotes it on, the index figure is 62 tok/s (community run) at 4-bit. The best measured figure is 61.6 tok/s on an M5 Ultra 256 GB (community run, 4-bit). Every Mac’s figure is in the table on this page with its provenance: published, community run or estimated.

What is the cheapest Mac that runs Qwen 3.5 397B-A17B well?

Among Macs Apple sells new, the Mac Studio M5 Ultra (256 GB), $9,499 new (Apple list price, checked 2026-11-14), is the cheapest that reaches 15 tok/s with Qwen 3.5 397B-A17B, at 62 tok/s (community run).

Do I need to raise the GPU memory limit to run Qwen 3.5 397B-A17B?

On 1 of the 54 configurations, yes: Mac Studio M3 Ultra 256 GB (used). macOS lets the GPU use only part of unified memory by default (about 192 GB on the Mac Studio M3 Ultra (256 GB)), and Qwen 3.5 397B-A17B needs about 210 GB. Running sudo sysctl iogpu.wired_limit_mb=245760 in Terminal raises the limit on a 256 GB Mac until the next restart. From 256 GB up, every Mac Apple sells new holds it inside the default limit.

Which quantization of Qwen 3.5 397B-A17B should I use?

Q4_K_M (about 226 GB of weights) is the common 4-bit GGUF choice. Q8_0 (about 421 GB) is closer to full quality and needs a 512 GB Mac at the default GPU limit. Q3_K_M and Q2_K shrink it to about 195 and 135 GB at a quality cost. The index’s reference figure (62 tok/s on an M5 Ultra) is for its 4-bit build.

How much context can Qwen 3.5 397B-A17B use on a Mac?

Its own window is 256K tokens. At macOS’s default GPU limit, the Mac Studio M5 Ultra (256 GB) holds about 256K (the full window).

Cite this page

LLMCheck. (2026). Can I run Qwen 3.5 397B-A17B on a Mac? llmcheck.net. https://llmcheck.net/can-i-run/qwen-35-397b-a17b/ (updated ). CC BY 4.0.

How to cite LLMCheck →Data: CC BY 4.0Markdown version

Related

Figures from the LLMCheck index, data version 2026-11-18. Verdicts and context limits are estimates from each model’s catalog memory figure and config.json.