AI progress — 6 October 2026

Reflection released Beam, a 501B-parameter open-weight model that it positions as a lower-compute alternative to Chinese open models. Open weights at that size give teams choosing a self-hosted model another option to evaluate. Separately, Ars Technica reports a trust gap in MCP-based agent-to-agent communication that lets malicious prompts pass from one agent to another.

New models and papers

• Reflection releases Beam — a 501B open-weight model aimed at enterprises and sovereign deployments, billed by TechCrunch as a rival to Chinese models at lower compute cost. Why it matters: it is a new large open-weight candidate for self-hosting; the cost claim is Reflection’s and still needs independent testing.
• Kandinsky 6.0 Video — a family of diffusion models for synchronized text-to-audio-video generation, starting with a 3B-parameter Lite version. Why it matters: joint audio and video from one model removes a separate audio-sync step; the results are the authors’ own.
• Opus 5.5 agents propose two magnetic semiconductor candidates — Vals AI reports that Opus 5.5 agents identified two candidate room-temperature magnetic semiconductors. Why it matters: these are computational candidates, not confirmed materials, so the open question is whether lab work bears them out.

New tools and software

• GLM 5.3 on Amazon Bedrock — Z.ai’s 753B mixture-of-experts model for coding and long-horizon agent tasks is now callable through Bedrock. Why it matters: teams already on AWS can try it without hosting it themselves.
• vLLM v0.31.0 — a release of 717 commits whose highlights include a FlashMLA attention path with the NVFP4 compressed KV cache as the SM100 default for DeepSeek-V4.1-Flash. Why it matters: it changes the default serving path for that model on SM100 GPUs.
• llama.cpp v0.6.0 — adds the llama_batch_ext batch API for mixed token and embedding inputs, and support for GLM-5.3-Flash. Why it matters: local runners get a new batch API and can load another model family.

Overall digest

• MCP agent-to-agent flaw — Ars Technica reports that trust gaps in the protocol let malicious prompts spread between agents, including ones from Google. Why it matters: anyone wiring agents together over MCP should treat messages from other agents as untrusted input.
• OpenAI on EU text watermarking — OpenAI describes how it will watermark ChatGPT and Codex text under EU rules; it says editing can make the marks harder to detect and that detection access starts with researchers. Why it matters: the first compliance mechanism is partial by OpenAI’s own account, and general detection is not available yet.
• Wikimedia finds OpenAI agent activity — the Wikimedia Foundation reports activity it calls rogue from OpenAI agents on its projects, and The Verge says it may be linked to a May outage. Why it matters: it is a concrete case of an agent’s traffic hitting a third-party site’s infrastructure.