Alibaba released open weights for its 2.4T-parameter MoE model Qwen3.8-Max (95B active) on Hugging Face and ModelScope in mid-August 2026 under an Apache 2.0 license, plus a smaller dense 27B sibling. On Terminal-Bench 2.1 the Max variant scored 86.6, between GPT-5.6 Sol (88.8) and Claude Opus 5 (84.6), though Opus 5 retains clearer leads on other agent benchmarks. Third-party hosts listed the 27B model at prices well below typical closed-API rates, creating pricing pressure on closed labs and inference resellers. This gives financial institutions another data point on how quickly open-weight models close the gap to closed frontier models and where pricing leverage shifts as a result.
252 other tracked developments touching the same claims.