Alibaba’s Qwen team published open weights for Qwen3.8-27B on Hugging Face Friday under an Apache 2.0 license, delivering the second half of an August 3 commitment and immediately pushing the release to the top of Hacker News, where it accumulated 893 points within hours.

The drop includes a full-precision checkpoint and an official FP8 quantization weighing in at 30.9 GB, the latter committed roughly two days ahead of wider confirmation. The model is a 27-billion-parameter dense multimodal system built on a Hybrid Gated DeltaNet architecture, with a 262K context window locally and a hosted version expected to default to 1M tokens. Thinking mode is on by default and tunable via a reasoning_effort parameter.

Benchmark numbers are the reason it’s getting attention. Studio Global reports the FP8 variant scoring 73.0 on Terminal-Bench 2.1, 61.7 on SWE-Bench Pro, and 90.3 on LiveCodeBench v6, a roughly 15 percent Terminal-Bench improvement over the April 22 Qwen3.6-27B and, notably, a result that beats Alibaba’s own API-only Qwen3.7-Plus on agentic coding. An NVIDIA deployment engineering post dated August 12 clocked the FP8 checkpoint at over 4,000 tokens per second per GPU on a GB300 NVL72 rack.

The structural read is more interesting than the scorecard. Chinese-origin models now account for roughly 61 percent of token volume on OpenRouter by mid-2026; the U.S. share fell from about 70 percent to 30 percent over the preceding year. An open, Apache-licensed 27B that outperforms the vendor’s own paid API tier is the mechanism by which that shift compounds.

Qwen3.8-Max, also promised on August 3, hasn’t appeared on the Qwen organization yet.

Sources