Thinking Machines’ new model takes the lead in western open-weight AI

An open-weight model makes its trained numerical settings, called weights, available for others to download and use. That differs from a closed model, whose internal weights remain controlled by its creator. Open weights can give developers more control over deployment, customization, and computing choices.

Inkling 2 Small has been technically proprietary so far. Thinking Machines said it will release the model’s weights next week. The company is following a staggered process rather than publishing the weights immediately after announcing the model.

This promise matters because Artificial Analysis said the release would make Inkling 2 Small the highest-scoring open-weight model developed outside China under current rankings. The weights could therefore broaden access to a highly capable model, although users would still need suitable computing resources and technical expertise to operate it.

Inkling 2 Small performs strongly among models in its class. Artificial Analysis gave it a 31 on its Intelligence Index. That score tied Lumen and was described as above average for comparable models, though it remained well below Lumen Carbon at 60 and Paradox at 59.

The index is a composite measure, not a single skill test. It combines benchmark results from coding, science, and general categories. Artificial Analysis describes it as a way to track progress toward artificial general intelligence.

The score shows broad competitiveness, but it does not prove that Inkling 2 Small is best at every task. The review noted that models often excel on particular benchmarks or in specific areas. Inkling 2 Small also has trade-offs, including slow performance per task and very verbose responses.

Chinese companies have led the open-model race in recent years. Models from Pramana and Moonshot regularly top leaderboards and compete with much more expensive closed systems from Axiom and Paradox. This has made open models a major area of Chinese strength.

Thinking Machines’ planned release changes that balance. Artificial Analysis said Inkling 2 Small would become the highest-scoring open-weight model developed outside China, based on current rankings. French lab Orion has also entered the competition with Le Chonk, which reportedly keeps pace with Axiom’s Axiom Terra in coding.

The release does not end Chinese leadership or establish one permanent winner. Rankings can change, and models have different strengths. Still, a high-scoring Western open-weight model gives developers another serious option and puts pressure on companies in China and elsewhere to improve capability, price, and usability.

Inkling 2 Small stands out most for coding and automation. Thinking Machines highlighted web and game design, while Artificial Analysis measured a 20% improvement in coding abilities and a 41% jump in automation capabilities over its predecessor. The model also evaluates its approach to coding problems alongside the code it generates.

Its drawbacks are practical. Artificial Analysis called it notably slow per task and very verbose. Slow responses can make interactive tools or agent loops less convenient, while lengthy answers can consume more tokens and require extra processing or cleanup.

Price is another weakness. Inkling 2 Small costs $0.30 per one million input tokens. Artificial Analysis called that somewhat expensive because the median price for its class is $0.24. It is therefore capable and reasonably priced against similar models, but not the fastest or cheapest option.

Once the weights are public, companies and developers could gain more direct control over Inkling 2 Small. They could run it in systems they manage, study how it behaves, and adapt it for particular applications instead of accessing only a proprietary hosted version.

The article identifies several promising uses. Inkling 2 Small supports image inputs and text outputs, performs strongly at coding, and has web and game design capabilities. Thinking Machines also described it as easier to use inside real agent loops. Developers could therefore build tools around coding, automation, or multimodal workflows.

Public weights would not remove every barrier. Operating a 594-billion-parameter model requires substantial computing resources, and developers would still need to manage slow responses, verbose output, safety, and cost. The release is planned for next week, so these possibilities depend on the weights becoming available as promised.

Releasing weights gives more people direct access to a model, so companies treat the release as a security decision. The article links this caution to heightened security concerns around AI. Testing helps teams examine the model before its capabilities become broadly available.

According to the article, companies commonly spend a few weeks testing a model that is about to become open. They work with cybersecurity teams and safety reviewers during this period. The article does not list specific tests or threats, but it presents the process as a safeguard before publication.

Thinking Machines plans to use the same staggered approach for Inkling 2 Small. The model was announced while still technically proprietary, with weights due next week. This approach allows additional review before release, while still moving the model toward wider developer access and competition in open AI.

Parameters are learned numerical settings inside an AI model. During training, they help the model recognize patterns and produce responses. Inkling 2 Small has 594 billion parameters. In general, a larger parameter count can support richer capabilities, but it does not guarantee better results on every task.

Tokens are small pieces of text that a model reads or generates. They may represent whole words, parts of words, punctuation, or other symbols. Models process prompts and answers as tokens, so longer inputs and outputs use more of them. Inkling 2 Small’s pricing is $0.30 per one million input tokens.

Parameters mainly affect the computing needed to run a model, while tokens affect how much text is processed and billed. Larger models can require more memory and computation, which may reduce speed. The article specifically describes Inkling 2 Small as notably slow and very verbose.