[AINews] Much ado about Open Weights

The open-weights debate hit peak theater this week. NVIDIA and Microsoft signed an open models letter, the internet turned it into a meme farm, OpenAI reportedly waffled before signing, and Anthropic sat it out entirely — later publishing a statement insisting it has never actually pushed for banning open-weight models. Lots of noise, very little movement, and mostly from a handful of players who will decide the real outcome regardless of who signs what.

While that argument consumed timelines, Moonshot AI just released Kimi K3. This is a 2.8 trillion parameter mixture-of-experts model with 104 billion active parameters, 896 experts total with 16 active per token, a 1 million token context window, and native visual understanding built in from scratch. Multiple independent evaluations reportedly show it beating Claude Opus 4.8, which would make it the strongest open-weights model available right now. On Agent Arena it's said to rank number one among open-weight models with a 9.75% net improvement, and Cognition called it the first open-source model they've tested that approaches frontier-level performance on their FrontierCode 1.1 benchmark, scoring 58.2% with a 63.6% pass rate.

Moonshot didn't just drop weights and walk away. Alongside K3, the company open-sourced FlashKDA (its attention kernels), MoonEP (a MoE communication library), and AgentENV (agent environment infrastructure) — effectively handing out the plumbing needed to actually train and serve models at this scale, not just the finished checkpoint. Day-zero availability across vLLM, Baseten, Modal, Together, Cursor, Cognition, Ollama Cloud, and Dell's Enterprise Hub turned the release into something closer to a supply-chain event than a typical research drop.

But