hn.today

Inkling 2 Mini

thinkingmachines.ai1376 points280 comments
Screenshot of Inkling 2 Mini

Inkling 2 Mini is an open-weight model designed to bring strong reasoning and agentic coding performance to laptop-scale hardware. Its 26.4 billion total parameters use a mixture-of-experts design, with 3.4 billion active per pass; the model repeatedly runs its middle layers three times for each token, spending computation on internal reasoning rather than producing longer chains of thought. It leads similarly sized open models on several coding tests, including DeepSWE v1.1 at 48.2% and SWE-bench Pro v2 Hard at 77.9%, and performs strongly in search and tool-use tasks. Its gains are not universal: repeated passes improve long-context reasoning, but do not add stored factual knowledge, so knowledge-heavy tests show little improvement. Search and tools help compensate for that limitation.

The release also presents looping as a way to improve reasoning without necessarily sacrificing oversight. In experiments with a variable-loop research model, increasing passes extended the no-chain-of-thought reasoning horizon from 16 to 37 seconds and improved serial-reasoning performance through ten loops, even though training used fewer. Written reasoning became shorter but remained legible, and tests found no increase in agentic sabotage as loops were added. Most weights use NVFP4 quantization, with the final four layers in MXFP8, keeping memory use near 20 GB for local use. The model is available with open weights and for inference and customization through Tinker; a demonstration showed it building a playable flight simulator offline on a MacBook Pro.

Read on thinkingmachines.ai280 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in AI

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.