Alibaba's answer to Meta's local model landed four days later — and this one can see
Alibaba's Qwen team published open weights for Qwen3.8-27B on 14 August under Apache 2.0. It is a dense 27-billion-parameter model, not a mixture-of-experts, and it is natively multimodal: it reads images and video as well as text. The model card lists a 262,144-token context window extensible to a million, and scores including 73.0 on Terminal Bench 2.1, 61.7 on SWE-bench Pro and 84.3 on OSWorld-Verified for computer use. Press reports put the memory requirement at roughly 24GB at usual precision — an RTX 4090 — with a community 4-bit build reported at about 17GB. Meta released Muse Glimmer, a 30-billion-parameter Apache 2.0 model with much the same pitch, on 10 August; it is text-only.
Why it mattersThe interesting number here is not the benchmark, it is the memory. Both of the last two open frontier-adjacent releases were sized deliberately to fit on one graphics card someone already owns, under a licence that lets them sell what they build. That is a different strategy from the trillion-parameter open releases of the past two weeks, which are open in name and unrunnable in practice for almost everyone. If a model that sees images and drives a computer now fits in 24GB, the set of things you no longer need to send to anyone's server just got considerably larger — for a school, a clinic, or a company with rules about where its data may go.
Read in the app — free, in 9 languages
Related stories
Rumour: the anonymous model that just topped a coding benchmark, for free, is said to be Zhipu's unreleased flagship2026-08-21DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21Show the robot once — three to twelve seconds — and it gets the job right 59 times out of 100 with no training at all
2026-08-21