Report: Qwen's 125B/6B open MoE, Flash-Next, previews Qwen 4
Weights aren't public yet and no benchmarks — but a 6B-active MoE you can self-host is a big deal. Plus: GPT-5.6 in Kiro, Copilot in Slack and Teams.
Copy markdown
125B total, ~6B active, open weights
Qwen 3.8-Flash-Next is reportedly a 125B-parameter mixture-of-experts that fires only ~6B parameters per token — the runtime footprint of a small model at a far larger capacity — with open weights expected on Hugging Face and ModelScope. The specs come from a ModelScope page that briefly went live and was pulled; there are no official benchmarks or license terms yet, so treat the "Sonnet-class coding" chatter as rumor until it actually ships.
GPT-5.6 lands in Kiro's spec-driven flow
OpenAI added the GPT-5.6 family — Sol, Terra, and Luna — to Kiro, its spec-driven coding agent that handles planning, building, review, and testing. OpenAI cites roughly an 82% cost reduction on Terminal-Bench 2.1 versus the previous generation, so the same agentic runs get materially cheaper.
Copilot agents move into Slack and Teams
GitHub Copilot is now in public preview inside Slack and Microsoft Teams — mention @GitHub in a channel, thread, or DM to start a collaborative agent session without leaving chat. It puts the Copilot CLI and app's agentic coding where your team already coordinates.
Elsewhere: a robot that learns from one video
Skild AI released S1, a robot foundation model that picks up a new multi-step task from a single human video with no fine-tuning — reportedly handling routines up to 10 minutes long and hitting 66% on tasks it has never seen. There's no download for builders yet, but it's a real step toward in-context learning for robotics.