The Radar

A 3B model trading punches with much larger ones is always worth a look. MIT licence keeps it genuinely open. The Spectrum-to-Signal post-training pipeline is the interesting methodological detail here.

Liquid AI is shipping search models that run on edge hardware and cover 11 languages. The ColBERT late-interaction variant is the interesting one for retrieval quality. Worth testing if you are building multilingual RAG pipelines outside of cloud-only setups.

Starting an agent with everything disabled except the bare essentials is a sensible default. Opt-in toolsets mean less surface area for things to go wrong. Small update, but the right instinct for production agent design.

No framework dependencies, MIT licence, and native WebMCP support out of the box. If you want agent UI without pulling in a large dependency tree, this is worth a look. Early days but the approach is clean.

Prompt injection and agent manipulation are real production problems. Lelu sits as an authorisation layer that catches agents acting outside their intended scope. Open source and early, but the category is increasingly important.

Agent sessions accumulate a lot of useful context that disappears the moment the chat closes. Callimachus indexes that history locally so you can search across it. Practical and immediately useful for anyone running coding agents daily.