Inkling 2 Small Offers Hosted Access, Weights Later
The 594 billion-parameter mixture-of-experts model is available now on Tinker, with open weights planned after safety testing.
Updated on Nov. 6, 2026 in Artificial Intelligence
Thinking Machines Lab released Inkling 2 Small on November 5, 2026, with closed weights on its Tinker platform. The model scored 31 on the Artificial Analysis Intelligence Index (v5.0) and is the second-generation release of the Inkling family; its weights are not yet public, and the company plans to release them under Apache 2.0 after external safety testing. The release is shipped as a hosted model, while the open-weights release remains announced rather than available.
Why it matters
The release shows a lab pairing a large mixture-of-experts model with a hosted-first launch and a delayed open-weights plan, a combination that lets it test capability and safety before weights circulate. It follows a rapid cadence within the family, with a 975 billion-parameter first Inkling model in July 2026 and a 276 billion-parameter Inkling-Small later that month.
The evidence is a third-party index score and a community design leaderboard, with Inkling 2 Small scoring 31 on the Artificial Analysis Intelligence Index (v5.0) and ranking 4th overall on DesignArena. The model uses 594 billion total parameters with 17 billion active per request, down from 975 billion in the first Inkling model. Whether these figures hold outside those two benchmarks has not been reported.
The players
Thinking Machines Lab
An AI research lab founded in February 2025 that develops the Inkling model family and serves it through the Tinker platform.
Tinker
The hosted platform through which Thinking Machines Lab serves Inkling 2 Small, charging $0.30 per million input tokens and $1.00 per million output tokens.
Artificial Analysis
An independent benchmarking firm whose Intelligence Index (v5.0) scored Inkling 2 Small at 31.
DesignArena
A community leaderboard that ranked Inkling 2 Small 4th overall and first among open-weights models.
The details
Inkling 2 Small uses a mixture-of-experts architecture, meaning it contains many specialized sub-networks called experts and routes each request to only a small subset of them. This is why its 594 billion total parameters translate to 17 billion active parameters per request. Training used a large integrated reinforcement-learning run, a method where the model learns from reward signals, that lets it use visual feedback to adjust its work. The model accepts image inputs and supports variable thinking effort, letting users trade response depth against speed.
Timeline
February 2025: Thinking Machines Lab launched with a $2 billion seed round at a $12 billion valuation.
July 15, 2026: The first Inkling model arrived.
July 30, 2026: Inkling-Small followed.
November 5, 2026: Inkling 2 Small was released on Tinker with closed weights.
The Tech Race
Inkling 2 Small follows the July 2026 Inkling and Inkling-Small releases, which were 975 billion and 276 billion total parameters respectively, and this model's 594 billion total parameters sits between them. Its hosted-first, open-weights-later sequence places it within a field where labs increasingly stage capability and safety review before releasing weights.
Inkling 2 Small is available now through the Tinker platform, which charges $0.30 per million input tokens and $1.00 per million output tokens. Developers who want to run the model themselves will have to wait for the open-weights release, for which no date has been announced.
The takeaway
The model's open-weights release is the milestone to watch, since it depends on external safety testing and has no announced date. Whether the 31 Intelligence Index score and the 4th-place DesignArena ranking hold up under further independent evaluation will show how far the mixture-of-experts design scales in practice.
What happens next
Thinking Machines plans to release the Inkling 2 Small weights under an Apache 2.0 license after external safety testing, but no date for that release has been given.
Further reading
For broader context on how these models are being built and deployed, see the Artificial Intelligence section.
