Qwen3.8-27B is available on @ollama ！Any harness you run, the model always deliv

Qwen, Alibaba's model team, announced on August 14, 2026 that Qwen3.8-27B can be run through Ollama, an app that downloads open models and serves them from your own machine. According to its model card, Qwen3.8-27B is a 27-billion-parameter dense model with open weights under the Apache 2.0 license that understands text, images and video, published at 16-bit precision. Ollama's library page, as read in November, lists the 27b version as an 18 GB download with a 256K-token context window, and the 27b tag shows that the file stores the weights in Q4_K_M, a four-bit compression format, rather than at the 16-bit precision Qwen published.

Qwen's post says the model always delivers in any harness, the agent program, such as a coding assistant, that gives a model its instructions and tools. The post offers no evidence for that. What the Ollama page does show is the practical connection: it lists commands such as ollama launch claude --model qwen3.8 that start Claude Code, OpenCode, Hermes Agent or OpenClaw with this model instead of their usual one.

The benchmark scores on Qwen's model card are not reported for this four-bit file, so they do not settle how the download performs. One independent benchmark (covered here) found that a 17 GB Q4_K_M build from Unsloth matched the full-precision model on the Terminal-Bench 2.1 coding test. That run used llama.cpp and a different file rather than Ollama's, the files were later replaced, and one test does not establish equal quality across tasks. The Ollama details above reflect the page in November and may differ from what was listed on the day of the post.