Edge Models, Leaked Source, and Post-Training at Scale

Tools

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

A 350M parameter model outperforming 700M+ models on several benchmarks is a real signal for edge deployment. The hybrid LIV architecture sidesteps the KV cache problem that makes transformers expensive on constrained hardware, making this worth evaluating if you're building for mobile or embedded.

TRL v1.0: Hugging Face's Post-Training Library Hits Production-Ready Status

TRL now covers 75+ post-training methods including PPO, DPO, and GRPO in one library, and v1.0 signals it's stable enough to build on. If you're doing any RLVR or preference optimization work locally, this is now the reference implementation to standardize on.

Anima Architecture: Open-Source Persistent Memory Framework Shows 59-Point Reasoning Improvement

The tiered memory loading approach using Notion and MCP is a practical alternative to vector database retrieval noise, and the 59-point gap on a cognitive battery is a concrete benchmark worth scrutinizing. Take the methodology with some skepticism since it's self-reported, but the architectural pattern of prioritized context loading is sound and replicable.

Analysis

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

The accidental .map file exposure gave a rare look at how a production AI coding agent is actually architected, including fake tool injection to poison training data scrapers. Practically useful for anyone building agent tooling, and the anti-distillation mechanics are worth understanding if you're in that space.

Claude Code Architecture Deep Dive: Five Layers, AsyncGenerator Loop, 100+ Tools

This is the more technical companion to the leak coverage, mapping out the full five-layer architecture including the QueryEngine, permission hooks, and plugin marketplace. If you build agent frameworks, the streaming AsyncGenerator pattern and permission model here are worth stealing directly.