LFM2.5-350M Doubles Down on Edge AI While Open Tooling Matures

Tools

LFM2.5-350M: Liquid AI Scales Training to 28T Tokens on a 350M Parameter Model

Liquid AI extended pre-training from 10T to 28T tokens and added large-scale RL, and the benchmark gains are hard to ignore — IFBench jumped from 18 to 40, tool use nearly doubled, and data extraction went from 11 to 32. If you're building pipelines that need fast, cheap inference at the edge or on-device, this is the most capable sub-400M model available right now. liquid.ai

TRL v1.0: Hugging Face's Post-Training Library Commits to Stability

After six years and 3 million monthly downloads, TRL finally ships a v1.0 with a hard stable/experimental split — SFT, DPO, GRPO, and RLOO are guaranteed stable, while 75+ experimental methods can still move fast. The intentional code duplication design choice is counterintuitive but makes sense given how quickly post-training research shifts. lilting.ch

Analysis

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

The accidental .map file leak keeps giving — Anthropic injected fake tool definitions to poison competitor training data, and there's a server-side summarization trick that strips reasoning chains from anyone recording API traffic. Whether you're building agents or just trying to understand where the industry is heading on IP protection, this is worth reading carefully. alex000kim.com

Claude Code Architecture Deep Dive: 1,884 TypeScript Files and a Full Agent Runtime

This is the most thorough technical breakdown of the leaked Claude Code source, covering the five-layer architecture, the AsyncGenerator loop that drives tool calls, and a permission model that treats humans as the final authority at every layer. If you're designing an agent framework, the QueryEngine pattern and how they handle streaming plus tool recursion is directly applicable. redreamality.com