AI News — September 22, 2026

$ cat top-story.md

▸ Kev Is a Self-Hostable, Open-Source Reproduction of Jev's "Decision Model" Trick

Days after TypeSafe AI's closed "Jev" made waves for returning typed decisions instead of text, developer Jared Palmer open-sourced Kev — a 0.8B/4B/9B family built on Qwen3.5 that runs on CUDA, ROCm or Apple Silicon and speaks the same API as TypeSafe's SDK. Why it matters: you can self-host the yes/no/rating "decision model" pattern for structured agent tool-calling instead of paying for the hosted version. // GitHub

▸ MiMo-V2.6-Flash Gives Self-Hosters a Lighter Alternative to Xiaomi's New Flagship

Alongside the Pro, Xiaomi released MiMo-V2.6-Flash — the same architecture scaled to 309B total/15B active parameters, MIT-licensed, priced at $0.435/$0.87 per million input/output tokens. Why it matters: a near-frontier open-weight model sized for single- or dual-GPU self-hosting, not just data-center clusters. // OrcaRouter

▸ Cua Reaches "Computer-Use 2.0" for Cross-OS Agent Fleets

Cua's latest release adds open-source drivers, cross-OS fleet management and benchmarks for training and evaluating computer-use agents, and is trending on GitHub with 600+ stars gained in a day. Why it matters: a reusable foundation if you're building or evaluating agents that need to click around a real desktop, not just call APIs. // GitHub

▸ ai-memory Gives Coding-Agent CLIs Long-Term Memory and Cross-Vendor Handoff

A Rust tool that persists long-term memory for coding-agent CLIs and lets you hand off context between different agent vendors instead of starting from scratch every session. Why it matters: solves the recurring pain of losing project context when switching between Claude Code, Codex and friends. // GitHub

▸ agent-native Is a New Framework for Building Agentic Apps

Builder.io's new TypeScript framework for building agent-native applications hit GitHub's daily trending list with over 600 stars gained in a day. Why it matters: worth a look if you're building your own agent UI/backend rather than assembling one from LangChain-style pieces. // GitHub

▸ Claude Code 2.1.278 Moves Auto Mode's Classifier to the Server, Off Your Bill

Auto mode now defaults to a server-side classifier for Claude API, Enterprise, Bedrock, Vertex and Foundry users, which doesn't charge for the classifier overhead; a new "Auto mode server" row in /status shows whether it's active. Why it matters: one less line item on your token bill if you rely on Auto mode to route requests. // Claude Code changelog

▸ Stanford's Paper2Agent Turns Any Research Paper Into a Callable MCP Server

Published in Nature, Paper2Agent uses multiple agents to analyze a paper's manuscript, code and datasets and build a tested MCP server exposing its methods — it stood up an AlphaGenome agent in about 45 minutes for roughly $14 in compute, beating the Biomni biomedical agent on genetics questions. Why it matters: a blueprint for turning any well-documented codebase or paper into an MCP tool your own agents can call. // Nature

▸ book-to-skill Turns Any Technical Book PDF Into a Cross-Agent Skill

Point it at a PDF or folder of docs and it distills the content into a structured skill — frameworks, decision rules, per-chapter files — usable from Claude Code, Copilot CLI, Amp or Hermes, claiming 24–51x fewer tokens than dumping the book into context. Why it matters: a practical pattern for building your own reference skills instead of re-explaining the same docs every session. // GitHub

▸ vLLM Details Prefill-Decode Disaggregated Serving for a 2.4T-Parameter Qwen Model

A new vLLM post walks through tuning prefill/decode-disaggregated serving for Qwen3.8-2.4T, reaching 5,000 tokens/sec per GPU in high-throughput mode or 180 tokens/sec/user for low latency, with reproducible topology and KV-cache recipes. Why it matters: the memory and topology tuning methodology applies just as well to smaller self-hosted deployments. // vLLM Blog

▸ Docling Is Trending Again as a Document-Prep Layer for RAG Pipelines

IBM's document-conversion library — which turns PDFs, tables and scanned documents into clean structured text — gained 130+ stars today and remains one of GitHub's most-starred RAG-adjacent tools at 67,500+. Why it matters: a solid, actively maintained building block for the document-ingestion step of a self-hosted RAG pipeline. // GitHub

▸ Fireship — "An Ex-OpenAI Researcher Just Deleted Language From the LLM"

Fireship's take on TypeSafe AI's Jev and the community reproductions (like Kev, above) that followed within days — a fast, technical explainer of why "decision models" that skip text generation are catching on for agent tool-calling. Why it matters: the clearest few minutes yet on why typed-output models matter for anyone building agents. // Fireship (YouTube)

▸ Linear: "AI Coding Has Made CI a Bottleneck, So We Reworked Ours to Keep Up"

Linear's engineering team details how AI-generated code nearly quadrupled their test suite, and how they cut PR wait times back down — new CI runners, the tsgo native TypeScript compiler, reworked Vitest sharding and trimmed critical-path jobs. Why it matters: a concrete playbook for the CI slowdown that hits any team once coding agents start producing most of the diff. // Linear, via Hacker News (172 pts)

▸ Amazon Blocks Meta's AI Shopping Agent, Citing Unauthorized Access

Amazon started serving an error page to Meta's "Muse" shopping agent, citing its Conditions of Use — a sign that e-commerce platforms are moving to lock out agentic browsing they don't control. Why it matters: a preview of the access battles ahead for anyone building agents that need to act on third-party sites. // TechCrunch