What's Actually Running at the Edge Right Now

Tools

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

Trained on 28T tokens with large-scale RL, this 350M model outperforms 700M+ rivals on benchmarks that actually matter for production use: structured output, function calling, and data extraction. If you're running inference on edge hardware or paying per token at scale, this is the model to benchmark against your current stack.

TRL v1.0 Ships: 75+ Post-Training Methods, Production-Ready Stability

TRL graduating to v1.0 signals that post-training is no longer just a research activity -- it's infrastructure. With GRPO, DPO, PPO, and verifier-based RL all under one roof, this is the fastest path to fine-tuning open models for specific tasks without building your own training loop.

Analysis

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

The accidental .map file drop revealed that Claude Code injects fake tools into API traffic to poison training data scraped by competitors, and uses server-side summarization to strip reasoning chains from recorded sessions. Whether or not you use Claude Code, these are real techniques worth understanding if you're building agent tooling or thinking about IP protection.

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

The leaked source reveals a full platform runtime with a multi-agent execution engine, plugin marketplace, cross-session memory, and a permission model that gates every tool call through human approval. If you're designing your own agent framework, the QueryEngine async generator loop alone is worth studying.

LFM2.5-350M's Hybrid LIV Architecture Solves the KV Cache Problem for Edge Inference

The model swaps most attention blocks for Linear Input-Varying convolution layers that keep constant-state memory instead of growing a KV cache with context length. That design choice is what makes it practical to run on CPUs and embedded hardware, not just the parameter count.