Skip to content
local-ai

Qwen3 30B-A3B

Alibaba · Text generation · MoE 30B-A3B · 131k context · Released 29 April 2025

Commercial use permitted Open weights Runs on CPU Apple Silicon

In a mixture-of-experts model, only a fraction of the parameters are used for any given token. Qwen3 30B-A3B activates 3B of its 30B parameters at a time, which means it runs at roughly the speed of a much smaller model while drawing on the knowledge of a larger one. The catch is memory: all 30B parameters must still be loaded, so it needs the memory of a 30B model even though it runs like a 3B one.

Strengths

  • Much faster generation than a dense model of similar quality
  • Strong reasoning for the speed, with optional thinking mode
  • Apache 2.0, long 128k context

Weaknesses

  • Needs the memory of a 30B model despite only 3B being active
  • A dense 32B can edge it on peak quality at similar memory
  • Thinking mode adds latency and token use

Hardware requirements

QuantisationApprox. VRAMNotes
Q4_K_M~18GBFits a 24GB card with some room for context
Q8_0~32GBNear-lossless, needs 36GB or more
FP16~61GBFull precision, server or multi-GPU territory

Also runs on CPU (slower). Optimised builds available for Apple Silicon.

What you'd need to run this

Roughly what a machine to run this would need, at up to three levels of quality. Memory is the deciding factor.

Minimum to run it

Q4_K_M · ~18GB needed

One 24GB GPU

NVIDIA Tesla P40

or a Mac or mini-PC with unified memory, if you prefer no discrete GPU, Mac mini M5 Pro .

32–64GB of system RAM alongside the card.

around £700–£1,100

What else 24GB runs →

For good quality

Q8_0 · ~32GB needed

64GB of unified memory

Mac mini M4 Pro

or a single high-memory workstation card, NVIDIA RTX 6000 Ada Generation .

Unified memory is shared with the model, so it is already counted above.

around £1,100–£2,000

See the Affordable Mac build build What else 64GB runs →

Best quality

FP16 · ~61GB needed

96GB of unified memory

AMD Ryzen AI Max+ 395 (Strix Halo)

or an 80GB-class data-centre card, which is usually rented by the hour, NVIDIA A100 80GB .

Unified memory is shared with the model, so it is already counted above.

around £1,100–£2,000

See the Quiet large-model box build What else 96GB runs →

Licence

Apache 2.0 — read the licence

Benchmarks

BenchmarkScoreSourceAs of
GPQA Diamond65.8 Qwen3 technical report (thinking mode) May 2025
AIME 202480.4 Qwen3 technical report (thinking mode) May 2025
AIME 202570.9 Qwen3 technical report (thinking mode) May 2025
LiveCodeBench v562.6 Qwen3 technical report (thinking mode) May 2025

How it compares

How this model’s reported scores sit against other models we cover, on the same benchmarks. This model is highlighted.

AIME 2025

higher is better

LiveCodeBench v5

higher is better

AIME 2024

higher is better

GPQA Diamond

higher is better

Qwen3 30B-A3B: common questions

What hardware do I need to run Qwen3 30B-A3B?
At its most compressed (Q4_K_M) it needs roughly 18GB of VRAM, and about 24GB for good quality. VRAM figures are approximate and depend on context length and settings.
Is Qwen3 30B-A3B free for commercial use?
Yes. Qwen3 30B-A3B is licensed under Apache 2.0, which permits commercial use with no meaningful conditions.
Can I run Qwen3 30B-A3B on Apple Silicon?
Yes. Qwen3 30B-A3B has builds optimised for Apple Silicon, through MLX or GGUF on a Mac.
Does Qwen3 30B-A3B run on CPU?
Yes, Qwen3 30B-A3B can run on the CPU, though generation is slower than on a GPU.
What is Qwen3 30B-A3B's context window?
Qwen3 30B-A3B has a context window of 131,072 tokens, about 131k.

Availability

Where to get quantised weights

Some of the best quantised weights are made by the community, not the model’s authors. Look this model up on these providers:

  • Unsloth GGUF (Dynamic 2.0, imatrix)

    Mixture-of-experts, so quant quality matters more at low bit-widths; dynamic quants help.

  • Bartowski GGUF Q2-Q8 (imatrix)

    A wide, reliable range of imatrix GGUF quants, typically Q2 through Q8.

  • MLX community MLX 4-bit and 8-bit

    MLX quants for Apple Silicon, usually 4-bit and 8-bit.

Recommended for

  • Users who want fast generation with strong quality on a 24GB card
  • Mac users who want responsive local inference at larger model quality
  • Interactive use where generation speed matters

Related models

Run it with

Related guides

Glossary

Our coverage

Catalogue entry last verified 30 July 2026. Specifications change; verify anything you are about to spend money on.