Qwen3.8-27B: a 27B model claiming 73 on Terminal-Bench and 84.3 on OSWorld
Alibaba dropped Qwen3.8-27B in FP8 and Hacker News put it at 733 points inside six hours. The number that makes people stop is 73.0 on Terminal Bench from a 27-billion-parameter model. For reference, GLM-5.3 made front-page news the same day for climbing to 28.3 on Terminal-Bench 3.0, which is a harder version of the same idea, but the point stands: this weight class was not supposed to be doing agentic terminal work at all a year ago.
The architecture is worth reading carefully because it is not a plain transformer. The layer pattern is sixteen repetitions of three Gated DeltaNet blocks followed by one gated attention block, each with its own FFN. Sixty-four layers, 5120 hidden. That is a hybrid linear-attention design, and the payoff is context: 262,144 tokens native, extensible to a million, at 27B. Cheap long context is the single thing agent workloads consume most, and a mostly-linear-attention model is how you get it without the KV cache eating your GPU.
It is also natively multimodal, which is where the other two numbers come from. 84.3 on OSWorld and 81.9 on AndroidWorld mean this thing drives a desktop and a phone. Those are computer-use benchmarks, not vision-QA benchmarks. A 27B open-weights model that you can serve on modest hardware and that operates a GUI is the configuration everyone building local agents has been waiting for, and it landed without a keynote.
SWE-bench Pro at 61.7 rounds it out. FP8 with fine-grained block-size-128 quantization, multi-token prediction in training, thinking mode on by default with adjustable depth and the option to turn it off. SGLang, vLLM and TokenSpeed are the recommended servers.
Set this against GLM-5.3 the same day and you get the shape of the week. Zhipu is holding weights for two weeks behind a safety review while claiming the strongest open coding model. Alibaba just put a multimodal agentic 27B on Hugging Face with no ceremony. Grab it at huggingface.co/Qwen/Qwen3.8-27B-FP8.
← Back to all articles
The architecture is worth reading carefully because it is not a plain transformer. The layer pattern is sixteen repetitions of three Gated DeltaNet blocks followed by one gated attention block, each with its own FFN. Sixty-four layers, 5120 hidden. That is a hybrid linear-attention design, and the payoff is context: 262,144 tokens native, extensible to a million, at 27B. Cheap long context is the single thing agent workloads consume most, and a mostly-linear-attention model is how you get it without the KV cache eating your GPU.
It is also natively multimodal, which is where the other two numbers come from. 84.3 on OSWorld and 81.9 on AndroidWorld mean this thing drives a desktop and a phone. Those are computer-use benchmarks, not vision-QA benchmarks. A 27B open-weights model that you can serve on modest hardware and that operates a GUI is the configuration everyone building local agents has been waiting for, and it landed without a keynote.
SWE-bench Pro at 61.7 rounds it out. FP8 with fine-grained block-size-128 quantization, multi-token prediction in training, thinking mode on by default with adjustable depth and the option to turn it off. SGLang, vLLM and TokenSpeed are the recommended servers.
Set this against GLM-5.3 the same day and you get the shape of the week. Zhipu is holding weights for two weeks behind a safety review while claiming the strongest open coding model. Alibaba just put a multimodal agentic 27B on Hugging Face with no ceremony. Grab it at huggingface.co/Qwen/Qwen3.8-27B-FP8.
Comments