Qwen's Flagship Broke Apache 2.0. Its 27B Sibling Just Restored It.

TL;DR

On August 14, 2026, Alibaba's Qwen team released Qwen3.8-27B, a 27.78-billion-parameter dense, native vision-language model with a 262,144-token context window extendable to 1 million. It is the local-hardware half of the Qwen3.8 generation, the piece that was still missing when the 2.4-trillion-parameter Qwen3.8-Max shipped on August 12 under a custom license that broke with two straight generations of Apache 2.0. The 27B keeps Apache 2.0 intact, jumps double digits over its predecessor on coding and computer-use benchmarks, and quantizes down to roughly 17GB, meaning it actually fits on a single GPU instead of demanding a cluster.

The Promise, Kept (Eventually)

Alibaba first teased the Qwen3.8 generation on July 19 with a single tweet and no benchmarks. On August 3, it promised open weights for both the 2.4T flagship and a 27B dense model "to meet you all" the following week. Only half of that showed up on schedule: Qwen3.8-Max landed August 12 with its parameter count, its shard count, and its new custom license, but the 27B was nowhere on the official Hugging Face org, only community reuploads pretending to be it. "Next week" stretched into eleven days, which by open-weight-lab standards counts as basically on time.

Apache 2.0 Survived This Generation After All

Qwen3.8-Max's license, a bespoke document with an attribution trigger above 100 million monthly users and a mandatory-negotiation clause above $50 million in aggregate revenue from inference resale or coding-assistant products, was the actual story when the flagship shipped. It broke a pattern that Qwen3.5 and Qwen3.6 had both held to. Qwen3.8-27B's LICENSE file is the plain Apache 2.0 text: use it, sell it, fork it, no revenue thresholds, no UI attribution requirement. For anyone planning to build a product on top of Qwen weights, the practical takeaway is that "the license got worse" wasn't a generation-wide shift, it was a Max-tier decision. The dense model most builders will actually run stayed exactly where it was.

The Numbers

Qwen built 3.8-27B on the architectural foundation of Qwen3.5, with 64 layers mixing Gated DeltaNet and gated attention and a 5,120 hidden dimension. Whatever the architecture changes bought, the benchmark table Alibaba published shows it landing hardest on long-horizon agentic work, not incremental gains:

The pattern holds elsewhere in the card: OSWorld-adjacent WebArena-Verified lands at 64.8 and AndroidWorld at 81.9, both without a prior-generation comparison published, and the vision side (math diagrams, chart reading, dense document parsing) posts similarly strong numbers. This is Alibaba's own table, not an independently reproduced one, but the direction, agentic and computer-use tasks improving faster than raw knowledge benchmarks, matches what every other lab has been optimizing toward this year.

Actually Runnable This Time

Qwen3.8-Max's smallest usable community quant was 397GB, a number that made the whole release, in Alibaba's own README's implication, a spectator sport for anyone without a rack of GPUs. Qwen3.8-27B does not have that problem. The official FP8 build lands around 28GB, and the Unsloth GGUF quantizations that appeared within a day of release go as low as 9GB, with the Q4_K_M build most people will actually want sitting at 17.1GB, comfortably inside a single RTX 4090 or 3090.

Quantization is the same trade you make recompressing a lossless photo into a smaller JPEG: shave enough bits per weight and the file shrinks dramatically while the output looks almost identical, until you push too far and the compression artifacts start showing up in the answers. At Q4 on a 27B model, that loss is close to invisible for most day-to-day use; at the 9GB end of the range, you are trading real capability for the ability to fit on smaller hardware.

Where It Sits

Alibaba's own claim, made directly in its release announcement, is that Qwen3.8-27B "outperforms Qwen3.7-Plus overall" despite Plus being a proprietary, API-only hosted model that only went generally available on June 1. If that claim holds up under independent testing, a self-hostable 27B beating a paid model from ten weeks earlier is a genuinely fast pace of depreciation for proprietary tiers. It also lands the same week as GLM-5.3, though the two releases are not really competing for the same job: Z.ai chased frontier cybersecurity and coding capability on a much larger model, while Qwen optimized a 27B dense model specifically to run on hardware a single developer already owns.

Key Takeaways

• Qwen3.8-27B shipped August 14, 2026, the dense open-weight companion Alibaba promised on August 3 alongside the 2.4T Qwen3.8-Max flagship
• It kept Apache 2.0 while Qwen3.8-Max, which shipped two days earlier, switched to a custom license with revenue-based restrictions
• Coding and computer-use benchmarks jumped hard over Qwen3.6-27B: SWE-bench Pro 53.5 to 61.7, Terminal-Bench 2.1 63.4 to 73.0, OSWorld-Verified 63.9 to 84.3, DeepSWE 1.1 more than tripling from 13.3 to 42.2
• It is actually self-hostable: Q4_K_M quantization lands at 17.1GB, fitting a single consumer GPU, versus the 2.4T flagship's 397GB minimum usable quant
• Alibaba claims it beats Qwen3.7-Plus, a proprietary hosted model that was itself only two and a half months old
• Native 262K context, extendable to 1M via YaRN, with full vision-language support for images and video baked in rather than bolted on

Sources: Qwen3.8-27B model card, repository file listing, Qwen3.8-27B LICENSE (Apache 2.0), Qwen (@Alibaba_Qwen) release announcement, The Decoder: Alibaba's Qwen team releases Qwen 3.8 models with open weights under the Apache 2.0 license, Unsloth Qwen3.8-27B-GGUF quantizations, Qwen3.8-2.4T-A95B model card