Cheaper to run and tune than full Inkling, but weaker on factual accuracy, so better suited to reasoning and coding work.

Kenning's take

What it does

Inkling-Small is a 276 billion parameter mixture-of-experts model from Thinking Machines Lab with 12 billion parameters active per token. The lab says it matches the full Inkling on many tests at about a quarter of the size, scoring higher on some reasoning and coding benchmarks but lower on knowledge and factual accuracy. Weights are on Hugging Face under Apache 2.0, and it can be fine-tuned and tried on Tinker.

• 276B total parameters, 12B active
• Text, image and audio input
• Up to 1M token context
• Adjustable thinking effort
• Lower cost and latency than Inkling

Who it suits. Developers who want an open multimodal model that is cheaper to run and fine-tune than Inkling.

Key facts

Category

AI models

Part of

Inkling

Made by

Thinking Machines Lab

Released

30 July 2026

How to get it

open weights on Hugging Face (Apache 2.0); fine-tuning and playground on Tinker

Works on

Hugging Face, Tinker, self-hosted

Free plan

Weights free to download under Apache 2.0

API price

Tinker: $1.16 per million prefill tokens, $2.88 per million sampled, $3.46 per million trained (list prices; limited-time 50% discount)

FAQ

Facts from the maker's own pages and the sources listed. Checked 2026-10-11.