Alibaba Releases Qwen3.8-27B Open Weights, a Dense Coding Model That Beats Claude Opus 4.6 Max on Two Benchmarks

Overview

Alibaba’s Qwen team released Qwen3.8-27B as open weights on Hugging Face on August 16, 2026, according to heise online. The 27-billion-parameter dense model is the smaller, on-premise-oriented checkpoint that Alibaba had signaled would follow its flagship Qwen3.8-Max release earlier in August, and it arrives with self-reported coding benchmark scores that beat Anthropic’s Claude Opus 4.6 Max on two of four headline tests, according to Qwen’s Hugging Face model card.

What We Know

Qwen3.8-27B is a dense, vision-capable model built on the architectural foundation of Qwen3.5, according to Qwen’s model card, which states the release “delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks.” It ships under an Apache 2.0 license with a native context window of 262,144 tokens, extensible up to 1,000,000 tokens, per the model card.

On Qwen’s own published benchmark table, Qwen3.8-27B scores 61.7 on SWE-bench Pro and 90.3 on LiveCodeBench v6 — both ahead of Claude Opus 4.6 Max’s 53.4 and 88.8 on the same tests, according to the model card. Opus 4.6 Max comes out ahead on the other two headline benchmarks Qwen published: 78.2 to Qwen3.8-27B’s 73.0 on Terminal Bench 2.1, and 91.3 to 89.2 on GPQA Diamond, per the same table. The model card also shows Qwen3.8-27B improving on its predecessors across all four tests, outscoring both Qwen3.6-27B and the closed-weight Qwen3.7-Plus, according to Qwen.

heise online technology writer Jan Mahn tested the model directly and reported it performs “equal to or better than Anthropic’s commercial Opus 4.6, which was released in February 2026,” according to heise online. In a hands-on coding test, Mahn had the model develop a REST API for an inventory management system with user management and role-based authorization, and reported that “the code compiled on the first attempt, and the ordered functions were present,” according to heise online.

For developers planning to run the model locally, heise online reported a range of hardware requirements depending on quantization: the full FP32 version needs at least 108 gigabytes of VRAM (120 gigabytes realistically once the key-value cache is included), the FP16 version requires cards with 64 gigabytes of VRAM, the Q5_K_M quantized variant fits into 22 gigabytes, and an NVFP4 variant optimized for Nvidia’s Blackwell-series compute units runs with as little as 16 gigabytes, according to heise online.

Qwen3.8-27B’s release configures a default reasoning-effort setting of “xhigh,” described on the model card as intended “for complex tasks demanding thorough analysis,” with “medium” and “low” options available for balancing accuracy, speed, and cost, according to Qwen.

What We Don’t Know

Qwen’s benchmark figures are self-reported and have not been independently replicated by a third-party evaluator. It is not yet clear how the model performs on coding benchmarks beyond the four Qwen chose to publish, or how its real-world agentic coding performance compares to Opus 4.6 Max outside curated test suites. Pricing and availability for Qwen3.8-27B through Alibaba Cloud’s hosted API were not detailed in the sources reviewed for this article.

Background

The Machine Herald previously reported that Alibaba made its flagship Qwen3.8-Max widely accessible on August 7, with open weights for both Qwen3.8-Max and “a smaller Qwen3.8-27B checkpoint positioned for on-premise GPU hardware” described at the time as still to come. Qwen3.8-27B’s August 16 release delivers on that on-premise checkpoint, giving developers a dense, single-machine-deployable alternative to the 2.4-trillion-parameter Max model.