ByteDance Seed and Tsinghua AIR open-sourced DAPO (Decoupled Clip and Dynamic Sampling Policy Optimization): algorithm, verl training stack, DAPO-Math-17k, and Qwen2.5-32B weights. Trained from scratch, DAPO-Qwen-32B hit 50% on AIME 2024 with about half the steps of DeepSeek-R1-Zero-Qwen-32B.
Amika’s Sept 20 write up pushes “Rigs”—reusable cloud VMs that keep repos, services, caches, and agents (Claude Code, Codex, OpenCode) across related tasks instead of one disposable sandbox per PR. Rigs can stay persistent, clone, or go ephemeral; secret and network controls are still labeled coming soon…
Alibaba shipped Qwen Image 2.1 weights on Sept 20: a 7B visual transformer with native 2K output, up to 10 reference images, and prompt driven RGB or RGBA transparency, already in Diffusers and ComfyUI. The catch is a research only license—commercial use needs a separate deal with Hangzhou Tongyi…
Tommaso Stocchi’s Sept 16 Agent Framework post argues many A2A specialists should become a SKILL.md plus typed MCP tools, with the parent advisor doing the reasoning. In a ski resort demo, model calls fell from 6–7 to 3 and mean latency from 15.5s to 6.3s, while total tokens rose about 22%…
Scalably’s Sept 20 report: two LoRA adapters trained on production agent traces improved recovery but regressed clean decisions, so both were rejected. Their takeaway for builders—keep fast memory and skills continuous, and change weights only behind hard gates…
Pirate Face launched a first party BitTorrent tracker and DHT discovery on Sept 20 so open weight model magnets can find peers after Hugging Face web seeds go dark. Its catalog lists 669k+ eligible repos, but only 4,891 witnessed checksums—and live seed counts stay thin on multi hundred GB models…
Spotted something we missed? Start a thread.