Edge Models, Post-Training at Scale, and What Leaked Claude Code Reveals

Tools

Liquid AI's LFM2.5-350M: A 350M Model Trained on 28T Tokens That Beats Models Twice Its Size

Liquid AI dropped a compact 350M parameter model that outperforms larger competitors on tool use, data extraction, and instruction following by moving away from pure transformer architecture toward hybrid LIV convolution blocks. If you're building edge applications or doing large-scale function calling, this is worth benchmarking against your current stack.

TRL v1.0: Hugging Face's Post-Training Library Hits Production-Grade Stability

TRL v1.0 ships with 75+ post-training methods including GRPO and DPO-style alignment techniques, and is now explicitly targeting production use cases rather than just research experiments. If you're doing any RLVR or preference optimization on open models, this is now the default starting point.

Analysis

Claude Code Source Leak: Anti-Distillation Fake Tools, Frustration Regexes, and Undercover Mode

Anthropic accidentally shipped readable source maps with the Claude Code npm package, exposing a multi-layered anti-distillation system that injects fake tools to poison competitor training data. The architectural decisions here are worth studying for anyone building production agent tooling, regardless of the drama around the leak itself.

Claude Code Architecture Deep Dive: Five Layers, AsyncGenerator Loop, and a Full Agent OS

A thorough teardown of the leaked Claude Code source across 1,884 TypeScript files reveals it's less of a chatbot wrapper and more of a platform runtime similar to VS Code's extension host. The async generator pattern driving the tool call loop is a clean design worth stealing for your own agent frameworks.