Timeline

Thinking Machines releases Inkling 2 Mini with open weights

A 26.4B-parameter mixture-of-experts model with 3.4B active that loops its middle layers three times; weights are on Hugging Face under Apache 2.0 from release day.

  • Models & capabilities
  • Open weights & ecosystem
  • Minor

Thinking Machines Lab released Inkling 2 Mini, a smaller companion to Inkling 2 Small. Small’s weights followed its announcement by a week; Mini’s Hugging Face repository was public, ungated and listed under Apache 2.0 on release day.

The post describes a mixture-of-experts model with 26.4 billion total and 3.4 billion active parameters, built as a “looped transformer” that runs its middle layers three times per token. Most weights are stored in NVFP4 for a footprint of about 20GB, which the company says suits consumer laptops.

Thinking Machines says Mini leads open-weight models under 100 billion parameters on benchmarks including DeepSWE v1.1 (48.2%) and the Vals Index v2.2 (46.5%). These are the company’s own reported figures, and its table notes that evaluation setups vary by model. The post also names weaknesses: Mini trails Qwen3.8-27B on multi-turn chat and constraint-heavy instructions, trails larger models on reasoning-heavy tests such as HLE-Diamond, and, the company says, looping adds reasoning rather than stored knowledge.

The company also argued that looping aids oversight, since the model can reason longer without writing its reasoning down. In a research variant released alongside, it reported a no-chain-of-thought time horizon rising from 16 seconds at one pass to 37 seconds at three, with chain of thought staying legible.