Latest / Models & Agents — Video Edition
Ep 149: A 250M-parameter model trained on 30B tokens now deploys in 60 MB with million-token…
Models & Agents A 250M-parameter model trained on 30B tokens now deploys in 60 MB with million-token retrieval from disk. What You Need to Know: A solo developer released SHADOW-250M, a heavily quantized LLM that keeps recent context in fp16 while compressing older tokens to 1 bit on disk. Nvidia published a linear-mapping technique that transfers KV caches between model sizes without full re-prefill. Simon Willison shipped llm 0.32.1 and llm-openrouter 0.7 with new tool support and compatibility fixes. Builders should test the tiny model for offline retrieval tasks and the KV transfer method…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.