Qwen 3.8: What Actually Shipped (and What It Means for Macs)

What Qwen 3.8 Actually Is

Qwen 3.8 shipped in August 2026 — but not the way the "Qwen 4" rumor mill predicted. The generation opened with Qwen3.8-Max, a hosted frontier model available only through Alibaba's API and chat products. Days later, the open-weight counterpart followed: Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter mixture-of-experts model with 95B active parameters. A Mac-class Qwen3.8-27B has been teased but had not shipped at the time of writing.

Here is the family at a glance:

The open-weight flagship posts serious numbers: 67.7 on SWE-Bench Pro and 92.6 on GPQA, putting it in the top tier of open models on published benchmarks. But its smallest published quantization is roughly 397GB — this is a server-class release, full stop.

The Open Weights — and the License Controversy

The bigger story than the parameter count is the license. Qwen3.8-2.4T-A95B is the first Qwen release that is not Apache 2.0. It ships under a custom license with a revenue threshold: broadly permissive for most users, but adding obligations for large commercial deployments built on the weights. Qwen built its open-source reputation on clean Apache 2.0 releases — Qwen 3.5 and 3.6 carry no such strings — so the move drew immediate criticism from the community.

Practically, the shift matters in two ways:

• For most individuals and small teams — the license is unlikely to bite. Running the model locally (if you had the hardware) or fine-tuning it for internal use falls well below any revenue gate.
• For companies shipping products — the terms deserve a real legal read before you build on Qwen 3.8 weights. If you want zero ambiguity, the Qwen3.6 line remains fully Apache 2.0, with no threshold and no field-of-use restrictions.

Alibaba is not alone here: this is part of a broader mid-2026 pattern of frontier-scale open releases pairing weights with revenue-gated custom licenses. For the license terms of every model in the catalog, see the LLMCheck leaderboard, which scores license openness as a component of every model's total.

Qwen3.8-27B: Teased, Not Shipped

The release Mac users actually care about is the teased Qwen3.8-27B, expected in August 2026. At the time of writing, neither its license nor its benchmarks had been published, so it is not yet in the LLMCheck index and there is nothing to score.

The open question is the license. If the 27B inherits Apache 2.0 like the Qwen3.6 line, it will be an automatic contender for the Mac #1 slot. If it inherits the new revenue-threshold license, the calculus changes for anyone building commercial products on local models. Until Alibaba publishes the terms, treat every "Qwen3.8-27B benchmark" you see as unconfirmed — we will add it to the leaderboard once there are official numbers to source.

What Mac Users Should Run Today

While Qwen 3.8 stays server-class, the Qwen3.6 generation remains the practical answer on Apple Silicon — and it is still excellent:

• Qwen3.6-27B — the Mac #1. A dense 27B under Apache 2.0 that scores 77.2% on SWE-bench Verified and carries an LLMCheck Score of 72, the highest of any Mac-runnable model in the index. Estimated throughput is ~40 tok/s on an M5 Max at Q4. See the full breakdown on the Qwen3.6-27B on M5 Max page.
• Qwen3.6-35B-A3B — the speed pick. A mixture-of-experts model that activates only 3B of its 35B parameters per token, so generation cost tracks a 3B model while quality tracks much larger. It scores 73.4% on SWE-bench and is Apache 2.0. Details on the Qwen3.6-35B-A3B on M4 Pro page.
• 8GB Mac? Try Bonsai 27B. Prism ML's native 1-bit/ternary derivative of Qwen3.6-27B (Apache 2.0, July 2026) squeezes the checkpoint to 3.9–5.9GB with vendor-reported ~90% quality retention — the first credible way to get Qwen3.6-class output on an 8GB machine.

New to local models? Start with the Ollama install guide, or wire Qwen3.6 into your editor with the local AI coding assistant guide. Not sure your hardware is up to it? The best LLM by Mac pages rank every model that fits your exact chip and RAM, and the best Macs for local LLMs page maps models to the machines that run them well.