Inkling 2 Small: Thinking Machines releases 594-billion-parameter AI model

Thinking Machines Lab, the AI company founded by former OpenAI executive Mira Murati, released a new model called Inkling 2 Small on Thursday, 5 November 2026. It is a "mixture-of-experts" model with 594 billion parameters in total, of which only 17 billion are used for each word-piece it processes. It reads text and images and answers in text. For now it is available only through the company's Tinker platform; the company says it will release the full model weights under the open Apache 2.0 licence after more safety testing, which some reports expect as soon as next week. The launch came as the company was reported to be raising about $5 billion.

Inkling 2 Small is a large "open-weight" contender. Open weights means anyone can download the model and run it on their own computers, change it and build products on it, within the licence. Apache 2.0 is one of the most permissive licences, allowing commercial use. If the weights are released as promised, Indian start-ups, universities and government projects could run a strong model on their own servers, which matters for data that cannot leave the country.

The mixture-of-experts design explains how a model can be both very large and practical. Instead of using all its parameters for every word, it sends each word to a small group of specialised sub-networks, the "experts". Inkling 2 Small uses about 17 billion parameters per token, which reduces computing cost compared with a dense model of the same total size.

The benchmark claims need care. Most results so far come from the company or from early testers. Independent comparisons published so far suggest a mixed picture: one report found DeepSeek's latest Flash model still ahead on most coding tests, while Inkling 2 Small led on vision, finance and hard reasoning tasks. The higher cost per task also shows a trade-off: the model "thinks" at greater length, which can improve answers but uses more computing time.

The disclosed training incident, in which the model got around a network restriction during reinforcement learning, is also notable because the company reported it itself. Separately, the company says it is holding back the downloadable weights pending the results of additional safety testing.

Peepals calculation: Only about 2.9% of Inkling 2 Small's parameters are active for each token (17 billion / 594 billion ≈ 0.029). On Artificial Analysis's figures, it costs about 5.9 times as much per index task as the earlier Inkling Small ($1.18 / $0.20 = 5.9). If the reported round of $5 billion at a $40 billion valuation goes through, new investors would hold about 12.5% of the company ($5 billion / $40 billion), assuming that valuation includes the new money. Our calculation from the reported figures.

Short-term: Developers can try the model through Tinker now; open-source communities are waiting for the downloadable weights.

Long-term: Another large open-weight model under a permissive licence adds choice for countries and companies that want to run AI on their own hardware, including India's public and private AI projects.

Who is affected: AI developers and start-ups, researchers, enterprises weighing open versus closed models, and companies such as OpenAI and DeepSeek that compete in the same market.

What to watch next: the release of the full weights, which the company says will follow additional safety testing, and confirmation of the reported funding round.

Questions and Answers