Alibaba's Qwen team has pushed its newest model generation into the open, and the headline is the sheer size of what it gave away. Qwen3.8-2.4T-A95B, a mixture-of-experts model with 2.4 trillion total parameters and 95 billion active per token, is now downloadable from Hugging Face in FP8 form. The team describes it as the first occasion a Max-class Qwen model has reached open release rather than staying behind a paid endpoint.
Landing alongside it is a far smaller counterpart aimed at a completely different audience. Qwen3.8-27B is a dense 27-billion-parameter model with a native vision encoder, able to read images and video rather than text alone. Between them the two checkpoints bracket the release: one built for organizations with serious cluster budgets, the other for teams that want capable inference on hardware they already own.
What Is Under the Hood
Both models inherit the architectural skeleton introduced with Qwen3.5. The flagship stacks 23 blocks that alternate three gated DeltaNet layers feeding mixture-of-experts modules against a single gated attention layer, routing each token through 10 experts plus one shared expert. The 27B model follows a similar alternating pattern but substitutes standard feed-forward networks for the expert routing.
Context length is identical on both: 262,144 tokens natively, extensible past a million. Each was trained with multi-token prediction across several steps, a technique that pays off in decoding speed at inference time. The FP8 packaging is not an afterthought either. Quantization runs at a block size of 128, and Qwen reports metrics that land within a hair of the unquantized originals.
Aimed Squarely at Agents
The pitch is less about answering hard questions and more about finishing long jobs. Qwen highlights stronger autonomous planning, better handling of environment feedback, and more reliable end-to-end completion of multi-step work. A tunable reasoning_effort parameter lets developers dial thinking depth per request, while preserve_thinking carries reasoning context forward across turns instead of discarding it.
The benchmark tables reflect that emphasis. Qwen positions the flagship against Opus 4.8, Fable 5, GPT-5.6 Sol and its own Qwen3.7-Max across agentic suites including Terminal Bench 2.1, SWE-bench Pro, DeepSWE 1.1, NL2Repo-Bench and FrontierSWE. Many of those evaluations were run inside the Claude Code harness, a rival vendor's tooling, with Qwen noting that its Max variant scores highest there on DeepSWE. The 27B model is measured against Opus 4.6 Max and Muse Glimmer-30B on a comparable spread, plus vision tests such as MathVision, CharXiv and SWE-MM.
Open Weights, With Caveats
The licensing is not uniform. The compact 27B ships under Apache 2.0, while the 2.4T flagship carries a custom license, a distinction that matters to anyone planning commercial deployment. Hosted options remain part of the plan: Qwen Cloud will offer a Qwen3.8-Max service and a managed 27B tier with a one-million-token default context window and built-in tools.
Community reception has been immediate, with both model cards climbing Hacker News and the flagship logging thousands of downloads within days. For developers, the practical question is now infrastructure rather than access. A 2.4-trillion-parameter checkpoint occupies roughly 2.5 terabytes even in FP8, which puts genuine local deployment beyond most budgets. The 27B model, at under 31 gigabytes, is the one most teams will actually run.
