Qwen 3 30B-A3B runs at 162 tok/s (162.2 as recorded) on a 256 GB M5 Ultra, a community run reported by bohemiansilicon.com (MLX, 4-bit, 512-token random prompt, 128 generated, mean of 3). It needs about 18 GB of unified memory. It is Alibaba's 30B mixture-of-experts model (3B active), released under the Apache 2.0 licence.
Want to run Qwen 3 30B-A3B faster — or step up to bigger models? Find your Mac →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.
| Metric | Value |
|---|---|
| Tokens per second | community162 tok/s |
| Provenance | community run · bohemiansilicon.com |
| Quantization | 4-bit |
| Engine | MLX |
| Prompt processing (prefill) | community7602 tok/s at 8K context |
| Context | 512-token random prompt, 128 generated, mean of 3 |
| Figure dated | 2026-11-11 |
| Minimum RAM | 18 GB |
| Parameters | 30B MoE (3B active) |
| Chip bandwidth | 1228 GB/s |
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 (18 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.
| Speed | Quant | Engine | Mac RAM | Context | Source | Note |
|---|---|---|---|---|---|---|
| sourced172 tok/s | 4-bit | llama.cpp | 256 GB | empty context, 128 generated | vsesmart.cz | Vse Smart review: Mac Studio M5 Ultra (80-core GPU, 256 GB), macOS 27.0.1, measured 24 Oct; llama.cpp with every layer on the GPU, build not stated. 4-bit; the review does not name the GGUF type. |
| community162.2 tok/s | 4-bit | MLX | 256 GB | 512-token random prompt, 128 generated, mean of 3 | bohemiansilicon.com | Bohemian Silicon, Mac Studio M5 Ultra (80-core GPU, 256 GB), macOS 27.2.1; mlx_lm.benchmark, mlx-lm 0.32.2 (MLX 0.32.4), mlx-community 4-bit. 138.0 tok/s with an 8K prompt. The post's second M5 Ultra unit measured within 0.35%. Published by Bohemian Silicon, which rents out this Mac. |
| sourced159 tok/s | 4-bit | MLX | 256 GB | empty context, 128 generated | vsesmart.cz | Vse Smart review: Mac Studio M5 Ultra (80-core GPU, 256 GB), macOS 27.0.1, measured 24 Oct; MLX 0.32 with mlx-lm, point version not stated. 4-bit MLX; the review does not name the repo. |
Measured figures only, each from its source’s own session; the index does not estimate long-context or prefill speed, so a missing cell is —. Each @ column is plain decode with that many tokens already in the prompt, from the same session as the row’s short-context figure: decode slows as the KV cache it reads grows. Prefill is prompt processing, the tokens per second a runtime reads a prompt at; with long agent contexts it sets how long you wait before the first token.
| Run | Decode (row) | @8K | @16K | Prefill |
|---|---|---|---|---|
| llama.cpp · 4-bit · 256 GB Mac · vsesmart.cz | sourced172 tok/s | — | — | sourced4535 tok/s |
| MLX · 4-bit · 256 GB Mac · bohemiansilicon.com | community162.2 tok/s | community138 tok/s | — | community7602 tok/s |
| MLX · 4-bit · 256 GB Mac · vsesmart.cz | sourced159 tok/s | — | sourced115 tok/s | sourced9870 tok/s |
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:
MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:
Downloads: Ollama page → · Hugging Face model card →
The 162 tok/s figure above came from MLX at 4-bit; 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.
Alibaba's Apache 2.0 30B MoE with 3B active. It fits a 24 GB Mac at 4-bit; tbreak's reviews measured 130.6 tok/s on an M5 Max and 62.0 on a Mac mini M6 (MLX, quantization not stated).
Apache 2.0 licence · 30B MoE · 128K context · LLMCheck capability 22/50 · see its leaderboard rank →
Estimated, not measured: the index’s minimum for Qwen 3 30B-A3B (18 GB — the weights at the quant that fits, plus ~2 GB) plus the KV cache for that many tokens, computed from its config.json. GQA on all 48 layers (4 kv heads x 128). 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 M5 Ultra Mac, which runs macOS 26 or later): “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).
Qwen 3 30B-A3B needs about 18 GB of unified memory. These Apple Silicon Macs have enough to run it. Apple builds the higher-memory configurations to order; Amazon does not stock them.
As an Amazon Associate, LLMCheck earns from qualifying purchases. These affiliate links cost you nothing extra and help keep the LLMCheck index free.
See how Qwen 3 30B-A3B stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard