Qwen3.8-27B Open Weights Ship — Single-GPU Companion Lands

Verdict: Qwen3.8-27B is no longer vaporware — it shipped, and we're moving it from pending to conditional. Open weights are live on Hugging Face and ModelScope under Apache 2.0, and the single-GPU companion to Qwen3.8-Max is real. The artifact exists and runs locally — but the benchmark gains are still Alibaba's own numbers, not independently reproduced, so it lands at conditional, not recommended.

What happened

On Aug 14, Alibaba's Qwen team announced(opens in new tab) that Qwen3.8-27B's open weights are live — "We promised open weights for Qwen3.8. Now, time to meet them!" The dense 27-billion-parameter model is the companion to the Qwen3.8-Max flagship, promised for the week of Aug 10 and overdue by four days.

• A native multimodal dense model. 27B parameters, not an MoE — built to run on a single GPU at ~15GB at 4-bit (or ~27GB FP8).
• Apache 2.0 license — confirmed by The Decoder, the same permissive terms Qwen used for prior open releases.
• 262K native context, extendable to 1M via YaRN, per Alibaba's official announcement.

The claimed benchmark leap

Alibaba's official model card(opens in new tab) numbers, versus the prior Qwen3.6-27B:

Qwen separately claims the 27B "outperforms Qwen3.7-Plus overall" — a different baseline from the Qwen3.6-27B comparison above, and notable for a 27B against a much larger Plus-tier model. These are vendor-claimed figures; the DeepSWE jump (13.3 → 42.2, a 3.2× gain) is exactly the kind of number that needs independent reproduction before anyone prices it into a pipeline.

Why it matters

The 27B is the Qwen3.8 artifact most developers can actually run. Qwen3.8-Max is a 2.4T-parameter MoE that demands datacenter hardware; the 27B fits on a single high-end GPU or an Apple Silicon machine. It sits in the same bracket as Meta's Muse Glimmer and Qwen's own Qwen3.6-27B — and if the coding benchmarks hold, it redefines what a local 27B can do.

That's why the verdict lands at conditional, not recommended: the "vaporware until it ships" reservation is resolved — it shipped under a permissive license — but every benchmark is still Alibaba-reported and the model is a day old.

This also completes the Qwen3.8 open-weight double-drop: the Max weights landed Aug 12, the 27B on Aug 14. Alibaba is now shipping open weights on a cadence that matters for every self-hosting team.

What changes for you

1. Grab the weights from the Qwen Hugging Face organization or ModelScope.
2. Run it locally — ~15GB at 4-bit (or ~27GB FP8) means a single 24GB consumer GPU (quantized) or Apple Silicon is enough to start.
3. Treat the benchmarks as unverified. Test against your own workload before swapping out Qwen3.6-27B in production.

FAQ

Is the 27B better than Qwen3.8-Max? They serve different brackets. The Max is a 2.4T-parameter MoE built for datacenter hardware; the 27B is the dense single-GPU model you can actually self-host.

Should I replace Qwen3.6-27B in production now? Not yet. The coding gains are Alibaba-reported — run the 27B against your own workload before migrating.

What hardware do I need? Roughly ~15GB at 4-bit (or ~27GB FP8). A single 24GB consumer GPU (quantized) or Apple Silicon is enough to start.