Moonshot released K3, a 2.8-trillion-parameter open-weight MoE with a 1M-token context window, currently topping Arena.ai’s Frontend Code leaderboard and trailing only Claude Fable 5 among benchmarked models.
Why it matters: Frontier-class open weights at $3/$15 per million tokens creates real pricing leverage for teams evaluating long-horizon coding and agent workloads.
The 975B-total/41B-active MoE is designed as a fine-tuning base through TML’s Tinker platform, with native text/image/audio support and controllable reasoning effort.
Why it matters: Another credible fine-tunable open-weight option expands the “build your own model” conversation for operators who want differentiation without closed-model lock-in.
The joint venture employs 100 engineers who embed in client organizations to build Claude-first solutions targeting CEO-level business transformations.
Why it matters: A $1.5B institutional bet that the trillion-dollar AI opportunity is in implementation, not models — signals where enterprise AI services are heading.
China’s Cyberspace Administration greenlighted Qwen to run Apple Intelligence across iOS/iPadOS/macOS/visionOS, opening 1B+ devices to on-device AI after a 16-month regulatory wait.
Why it matters: The Qwen/Apple integration — compressed to under 4GB for on-device inference — is a proof point for how AI stacks will increasingly mix frontier APIs with local models as costs fall.