Thinking Machines' Inkling 2 Small leads western open models in tests

Thinking Machines Lab has released Inkling 2 Small, and independent tester Artificial Analysis ranks it as the strongest open-weight AI model built outside China once its weights are published. The model scores 31 on the Artificial Analysis Intelligence Index, and Thinking Machines says it will release the weights next week.

This is an update to Unicorn API’s earlier overview of the family, Thinking Machines’ open Inkling models: what the 276B Inkling-Small offers. What changed since then: a second-generation small model now exists, it is larger than the first, and the first third-party scores for it are out. One caveat applies to the whole story. The “open-weight” ranking depends on a release that has not happened yet. ZDNET notes that the model is technically proprietary until the weights ship.

What Thinking Machines released and what Artificial Analysis found

Thinking Machines Lab is the startup founded by former OpenAI chief technology officer Mira Murati. Inkling 2 Small is its latest model, and ZDNET reports it on 2026-11-06. The model listing site AI/TLDR gives a release date of 2026-11-05.

The headline finding comes from Artificial Analysis, which runs a composite benchmark called the Intelligence Index. ZDNET describes the index as a test “for tracking progress toward artificial general intelligence” that combines benchmark scores across coding, science and general categories. Inkling 2 Small scores 31. That is above average for models in its class and tied with Gemini 3.8 Flash. Artificial Analysis states the ranking claim directly:

“[Thinking Machines’] plan to release its weights would make it the highest-scoring open weights model developed outside China based on current rankings.”

That quote comes from Artificial Analysis as cited by ZDNET. For scale, the top of the same index is held by closed models: Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59. Inkling 2 Small leads a specific category, western open-weight models. It does not challenge the overall frontier.

What Inkling 2 Small adds over Inkling-Small

The new model is much larger than its predecessor. Inkling 2 Small has 594 billion parameters. The original Inkling-Small, released on July 30, 2026, has 276 billion total parameters, of which 12 billion are active for each token. A parameter is one of the internal values a model learns during training. The sources do not give an active-parameter count or an architecture breakdown for Inkling 2 Small.

Inkling 2 Small accepts image inputs and produces only text. Thinking Machines highlights its web design and game design capabilities. The company also says it evaluated how the model approached coding problems during testing, alongside the code it produced. The company describes the outcome this way:

“The result is a model that is both stronger at coding and easier to use inside real agent loops.”

That is Thinking Machines’ own claim, as quoted by ZDNET. An “agent loop” is a setup where a model repeatedly plans, calls tools or writes code, checks the result and continues.

Artificial Analysis reports gains over the predecessor in its own testing. Automation capabilities rise 41% and coding abilities rise 20%. ZDNET adds a caution: most models do well in particular areas or on particular benchmarks, which makes it hard to name any single model the best coder overall.

Fewer hallucinations, more abstentions

The clearest change in Artificial Analysis’s numbers is in factual reliability. On its AA-Omniscience v1.1 test, the organisation reports that “the hallucination rate falls from Inkling Small’s 61% to 27%, while accuracy rises from 32% to 34% and the attempt rate falls from 77% to 53%”. A hallucination is a confident but wrong answer. Read together, these figures say the model now declines to answer far more often than it did before. Its rate of correct answers moves only slightly. For builders, a model that abstains more and fabricates less is often easier to put into production, but it can also leave more questions unanswered.

This fits the design goals Thinking Machines set for the first Inkling release, which was built to flag uncertainty rather than guess.

Inkling 2 Small specs at a glance

Spec	Inkling 2 Small	Inkling-Small (July 2026)
Parameters	594B	276B total, 12B active
Inputs	Text and images	Text, images and audio, listed for the Inkling family
Output	Text only	Text
Context window	1M tokens, per AI/TLDR	1M per AI/TLDR; 524K per ATLABYTE
Licence	Tinker Terms of Use, Apache-2.0 weights planned	Apache-2.0
Weights	Planned for next week	Published
AA Intelligence Index	31	Not given in the sources

The sources disagree on the original Inkling-Small’s context window. AI/TLDR lists 1M tokens, while ATLABYTE lists 524K. The context window is the amount of text a model can consider at once. The difference may reflect different deployment configurations, but the sources do not say.

Access, licence and the staggered weight release

The open-weight label depends on a promise. Thinking Machines says it will release the weights next week. ZDNET describes this staggered approach as increasingly customary. Given security concerns around AI, companies spend a few weeks testing a model with cybersecurity teams and safety reviewers before releasing the weights, instead of publishing them on day one. Mistral followed the same process with its Le Chonk model, covered earlier in Mistral releases Le Chonk, its 1-trillion-parameter open-weight model.

On licensing, AI/TLDR currently lists Inkling 2 Small under “Tinker Terms of Use (Apache-2.0 weights planned)”. Tinker is Thinking Machines’ platform for customising and fine-tuning models. The ZDNET report does not state a licence for the upcoming weights. Builders should treat Apache-2.0 as expected but unconfirmed until the release. The earlier Inkling and Inkling-Small weights were both published under Apache-2.0, a permissive licence that allows commercial use.

On cost, Artificial Analysis calls the model’s pricing “somewhat expensive” at $0.30 per million input tokens, compared with a median of $0.24 for its class. Its broader verdict is mixed:

“Inkling 2 Small is amongst the leading models in intelligence and reasonably priced when comparing to other models of similar price. It’s also notably slow per task and very verbose.”

That assessment is quoted by ZDNET.

Why a western open-weight leader matters

The context for this ranking is that Chinese labs have led open models for several years. Models from DeepSeek and Moonshot regularly top leaderboards and rival more expensive closed models from OpenAI and Anthropic, according to ZDNET. Since DeepSeek’s R1 performed similarly to OpenAI’s o1 in January 2025, Chinese open models have given enterprise customers reasons to rethink costly contracts with US labs.

When Thinking Machines launched the first Inkling in July, Reuters reported that the West’s open-source ecosystem lagged China’s. Reuters linked this partly to Meta’s shift to a proprietary approach after its Llama 4 release, and noted that businesses had flocked to Chinese models as alternatives to expensive closed ones. At that launch, Thinking Machines’ own benchmarks showed rival models keeping the edge overall, with Inkling competitive particularly on agent-related tasks. The company has framed the Inkling line as a base for customisation through Tinker rather than a model meant to top leaderboards.

Inkling 2 Small does not change that framing, but it does change the comparison. For teams that need open weights from a non-Chinese developer, whether for procurement, compliance or preference, there is now an option that an independent tester places at the top of that group.

What it means for builders, and the open questions

• Agent and coding work: Thinking Machines pitches the model for agent loops, and Artificial Analysis measures gains in automation and coding. Its slowness per task and verbosity mean longer runs and more output tokens, which matters for latency-sensitive agents and for total cost.
• Reliability: The drop in hallucination rate to 27% is the most practical improvement for retrieval and question-answering products. Expect the model to decline to answer more often.
• Modalities: Coverage of Inkling 2 Small mentions image input and text output. It does not mention audio input, which the original Inkling family supported. Teams relying on audio should check the release details.
• Self-hosting: At 594 billion parameters, running the model yourself will need substantial hardware. Without a published active-parameter count, inference cost is hard to estimate.

Several questions remain open. The weights are not public, so none of the scores can yet be reproduced on self-hosted copies. The licence is listed as planned rather than final. The evidence also rests on a single third-party evaluation, Artificial Analysis, reported through ZDNET. Further independent testing would confirm or qualify the ranking.

What to watch next

The immediate milestone is the weight release that Thinking Machines has promised for next week. Watch the licence that accompanies it, whether the published checkpoints match the model Artificial Analysis tested, and which hosting providers list it. The original Inkling was served by providers including Together AI, Fireworks AI, Modal, Databricks and Baseten. Also watch whether Chinese labs answer with new open releases that push the bar above an index score of 31, and whether other independent leaderboards agree with Artificial Analysis’s placement.