[UPDATE] Alibaba open-sources the Qwen3.8 line: a 27B multimodal dense model under Apache 2.0, plus a 2.4T MoE
Alibaba published Qwen3.8-27B, a 27-billion-parameter multimodal dense model with 262K native context under the permissive Apache 2.0 licence, alongside Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts system activating about 95bn parameters at inference. Reporting places both on HuggingFace with 17 quantised variants. It lands in the same week that OpenAI cut GPT-5.6 Luna about 80% and Anthropic priced Claude Opus 5 near half of Fable 5, and is the supply-side pressure those cuts are being attributed to.