Moonshot AI published Kimi Linear on Thursday: a 48B-parameter mixture-of-experts model that activates 3B parameters per token, carries a 1M-token context window, and was trained on 5.7 trillion tokens. The architecture interleaves three Kimi Delta Attention layers — a channel-wise-gated refinement of Gated DeltaNet — for every full multi-head latent attention layer. Moonshot says that 3:1 ratio cuts KV-cache demand by up to 75% and reports 1.84 ms versus 11.48 ms per output token against MLA at 1M tokens, a 6.3x decode advantage. Checkpoints and the KDA kernel are open-sourced.

The claim worth taking seriously is not the speed table but the quality comparison: the report argues this is the first hybrid linear attention design to beat full attention under fair comparison across short-context, long-context and RL-scaling regimes, with MMLU-Pro at 51.0 against MLA's 47.2 and RULER at 128K at 84.3 against 81.3. Linear attention's historic failure mode is exact recall — a fixed-size state cannot copy arbitrary earlier tokens — so the interesting question for builders is where that trade still bites, not whether the benchmark moved.

OpenAI released a research preview of gpt-oss-safeguard on Wednesday: open-weight reasoning models at 120B and 20B, fine-tuned from gpt-oss and published under Apache 2.0. Instead of a fixed classifier trained on labeled examples, the model takes two inputs — a developer-written policy and the content — and reasons its way to a label, which means policies can be changed without retraining. OpenAI is explicit about the trade-offs: a classifier trained on tens of thousands of labels can still outperform it, and it is compute-intensive enough that OpenAI runs smaller high-recall classifiers in front of its internal Safety Reasoner.

The next day, Chain-of-Thought Hijacking showed what happens when reasoning is the mechanism you are trusting. Wrapping a harmful request inside a long benign puzzle-solving trace — often five minutes of reasoning — defeated refusals 99% of the time on Gemini 2.5 Pro, 94% on ChatGPT o4 Mini and Claude 4 Sonnet, and 100% on Grok 3 Mini; activation probing suggests refusal depends on a low-dimensional safety signal that weakens as traces lengthen. Reasoning that reads a policy is also reasoning that is long enough to forget one.

OpenAI completed its recapitalization on Tuesday, converting the for-profit into a public benefit corporation called OpenAI Group PBC, controlled by the renamed OpenAI Foundation. The Foundation holds equity currently valued at about $130 billion, a 26% stake; current and former employees and investors hold 47%; Microsoft holds roughly 27%, valued at about $135 billion. The reworked Microsoft agreement extends Microsoft's IP rights through 2032 — now including post-AGI models — drops Microsoft's right of first refusal to be OpenAI's compute provider in exchange for an incremental $250 billion Azure purchase, and requires any AGI declaration to be verified by an independent expert panel.

The post from board chair Bret Taylor notes the deal closed after nearly a year of dialogue with the California and Delaware Attorneys General — the regulatory follow-on to the 2024 for-profit conversion that OpenAI abandoned in May 2025. Microsoft's own summary of the agreement carries a line developers should read twice: OpenAI is now able to release open weight models that meet requisite capability criteria. Read alongside the same week's model drops, Monday and Wednesday stop looking like generosity and start looking like contract terms.

MiniMax open-sourced M2 on Monday: a mixture-of-experts model with 230B total parameters and 10B activated per token, built for tool use rather than conversation. It carries roughly a 200K-token context window, and MiniMax lists API pricing at $0.30 per million input tokens and $1.20 per million output, which the company puts at about 8% of Claude Sonnet 4.5's price at roughly double the inference speed. Weights are on Hugging Face with day-0 serving paths in vLLM, SGLang and MLX.

Reported benchmarks put M2 at 69.4% on SWE-bench Verified and 46.3 on Terminal-Bench, with the top composite score among open-source models on Artificial Analysis' index. The number that decides whether you can use it is the 10B activation: 4-bit weights land near 130GB, so a 2x80GB node serves long agent loops at interactive speed. One operational catch — M2 is an interleaved-thinking model, and its thinking blocks must be preserved in message history or quality degrades.