Release2 sources
The Loop Is a Clock, Not a Library
DEV · By Puneet Nanda ·
The Brief · written from 2 sources
- Thinking Machines Lab released Inkling 2 Mini on November 20, 2026.
- It is a 26.4B-parameter open-weights model that the sources say runs locally on a laptop.
- It is a looped transformer whose middle layers run three times per token; on the lab's variable-loop research model, more passes helped serial lookups but not trivia.
Why it matters
Developers can run a coding-agent-capable open-weights model locally, and the looped design suggests extra compute per token helps multi-step reasoning more than factual recall.
Who’s affected
- Developers wanting to run coding agents locally
- Teams evaluating open-weights models
- #llm
- #open-weights
- #local-inference
- #looped-transformers
- #coding-agents
Another pass buys a step. It does not buy a fact. Inkling 2 Mini runs its middle layers three times per token. On the lab's variable-loop model, more passes helped a lookup chain and did not help trivia. That research model can change the loop count. Serial lookups kept improving…