GLM-5.2 goes huge, Baidu flattens the KV cache, Claude moves into Slack

GLM-5.2 is probably the most powerful text-only open weights LLM — Z.ai dropped a 753B MoE under MIT license with a 1M token context — text only, vision lives in the separate (closed) GLM-5V-Turbo. Past GLM releases have flattered to deceive on benchmarks vs Claude and GPT-4, so the only number that matters is the one you measure on your own workload. — simonwillison.net

Baidu's Unlimited OCR keeps the KV cache flat — a 3B MoE with Reference Sliding Window Attention that holds the KV cache constant as output grows — memory and latency stop scaling with document length. Scores 93.23 on OmniDocBench against DeepSeek's 87.01. If your pipeline chews on long PDFs, this is the kind of architectural trade-off worth benchmarking. — Asif Razzaq

Claude Tag: multiplayer, proactive, persistent agents in Slack — Anthropic's Slack integration lets Claude hold state across threads and coordinate with multiple users — closer to a coworker than a stateless API call. Useful if your team already lives in Slack; less so if you don't. Pricing details remain the obvious tell on how seriously they want enterprise adoption. — latent.space

Open SWE: LangChain's open-source async coding agent — plans, codes, tests, and opens PRs on GitHub autonomously, with a cloud-hosted runtime. The interesting question isn't whether it works on toy repos — it's the task completion rate on your actual backlog versus Devin, Cursor agents, or Claude Code. — langchain.com

Krea 2: SOTA open-weights 12B image model — open-weights image models at this quality tier remain rare. If Krea 2 holds its own against Flux and DALL-E 3 at 12B parameters, it sets a new efficiency bar for the field. Worth a head-to-head on your prompt set before swapping anything in production. — krea.ai

Haystack: open-source framework for production agents and RAG — deepset's 2.0 pitches modular components — swap LLM providers or retrieval backends without rewriting pipelines. Positioned against LangChain on production stability rather than feature surface area, which is the right axis if you've ever shipped a RAG demo and then tried to keep it alive. — doener

DATA

Engineering jobs are the most resilient, not the most threatened — SignalFire data shows engineers making up a larger share of new hires, not smaller, despite the layoff-narrative everyone keeps repeating. Useful counterweight if you're a founder modeling headcount or an engineer pricing your career risk. — Marina Temkin

AI researchers keep leaving Google for rivals — Jonas Adler and Alexander Pritzel are the latest to decamp for Anthropic, following Noam Shazeer and John Jumper out the door. Talent concentration is a reasonable proxy for which lab actually executes on the next capability jump — and increasingly it isn't the one with the most TPUs. — Amanda Silberling, Lucas Ropek

Norway imposes near-ban on AI in elementary schools — citing child development and learning-outcome concerns. Worth a bookmark if you're building ed-tech or sourcing training data from classrooms — if larger EU economies follow Norway's lead, the addressable market shifts overnight. — reuters.com