When 350M Parameters Beats the Field and Post-Training Gets a Stable Home

Tools

Liquid AI's LFM2.5-350M Beats Models Twice Its Size on Tool Use and Instruction Following

A 350M parameter model trained on 28T tokens now outperforms 800M+ models on benchmarks like BFCLv3 and IFBench. If you're running inference at the edge or at scale, this changes the cost calculus significantly.

TRL v1.0 Lands as the Go-To Post-Training Library for Production Use

TRL now covers 75+ post-training methods including PPO, DPO, and GRPO in a single stable library built for real production systems. This is the clearest signal yet that post-training is a first-class engineering discipline, not just research tinkering.

Analysis

The Claude Code Source Leak: Anti-Distillation Tricks, Frustration Regexes, and Undercover Mode

Anthropic accidentally shipped a readable source map with their Claude Code npm package, exposing fake tool injection designed to poison training data scrapers and a server-side summarization trick that strips reasoning chains from API traffic recordings. If you build on top of closed AI APIs, understanding these defensive patterns is worth your time.

Claude Code's Five-Layer Architecture Is Closer to VS Code Than a Chatbot Wrapper

The leaked source reveals Claude Code is a full platform runtime with a multi-agent loop, plugin marketplace, and cross-session memory system built on an async generator pattern. Builders designing their own agent frameworks will find the QueryEngine and tool permission model worth studying closely.

News

Hybrid LIV Architecture Lets LFM2.5-350M Skip the KV Cache Problem Entirely

Unlike pure Transformer models, LFM2.5-350M uses a hybrid backbone with Linear Input-Varying convolution blocks that maintain constant-state memory, avoiding quadratic KV cache scaling. This is the architectural reason the model runs efficiently on cheap CPUs and embedded hardware, not just marketing copy.