Thinking Machines releases Inkling 2 Small, scores 31 on AI Index

Thinking Machines Lab has released Inkling 2 Small, the newest member of its model family. It scored 31 on the Artificial Analysis Intelligence Index (v5.0).

The company calls itself committed to open weights. For now, though, the model ships as a proprietary product, with the weights promised later.

The launch landed on November 5, 2026. It is the lab’s second-generation release, arriving less than four months after the original Inkling lineup debuted in July.

What’s under the hood

Inkling 2 Small uses a Mixture-of-Experts architecture. It has 594 billion total parameters, but only 17 billion are active at any given time.

Each request gets routed to a small subset of the model’s internal “experts.” The result is a model that holds a lot of knowledge but only spends a fraction of its compute on any single query.

Usage through Thinking Machines’ Tinker platform costs $0.30 per 1 million input tokens and $1.00 per 1 million output tokens. The research findings describe that pricing as competitive with peers.

The model accepts image inputs and supports variable thinking effort. Users can dial up how hard it reasons on a problem, trading speed and cost for depth.

Thinking Machines is pitching it for agentic coding, vision, reasoning, and knowledge-work tasks. “Agentic” means the model can carry out multistep jobs on its own rather than simply answering one question at a time.

Training leaned on a large integrated reinforcement-learning run. That process lets the model use visual feedback to adjust its work in tasks like building websites and games.

On benchmarks, Inkling 2 Small ranked 4th overall on DesignArena. It also placed 1st among open-weights models on that leaderboard.

Its AA Intelligence Index score of 31 gives a single-number snapshot of general capability. The index is version 5.0, so the score reflects the latest edition of that test suite.

How it got here

Thinking Machines Lab launched in February 2025 with a $2 billion seed round at a $12 billion valuation.

The first Inkling model arrived on July 15, 2026. It was a larger system, with 975 billion total parameters and 41 billion active.

Inkling-Small followed on July 30, 2026, with 276 billion total parameters and 12 billion active. The new release sits between those two in size.

Inkling 2 Small is more than twice the total size of its namesake predecessor. Its active parameter count rose from 12 billion to 17 billion.

What this means

The most interesting tension is the weights question. Thinking Machines frames itself around an open-weights approach, yet Inkling 2 Small currently runs only as a closed model on Tinker.

The company plans to release the weights under an Apache 2.0 license after external safety testing. Apache 2.0 is a permissive license that generally lets anyone use, modify, and build commercial products on the software.

Until that release happens, developers can only access the model through Thinking Machines’ own platform. They can run inference and fine-tune it there, but they cannot download it and host it themselves.

No date has been attached to the weights release.

At $0.30 and $1.00 per million input and output tokens, the model targets teams that want strong coding and design capability without premium-tier bills.

The DesignArena results reinforce that pitch. A 4th-place overall finish, combined with the top spot among open-weights models, suggests the visual feedback training is paying off in front-end and design work.