Thinking Machines Lab Releases Inkling 2 Small, a 594B-Parameter Model

Thinking Machines Lab, the company led by former OpenAI CTO Mira Murati, launched a new AI model called Inkling 2 Small on November 5, 2026. This model is a Mixture-of-Experts with a total of 594 billion parameters and 17 billion active parameters per token. The model also has a vision encoder of 292 billion parameters.

The model currently runs closed on the Tinker platform. Full weights with Apache 2.0 license are planned for release next week after external safety testing. The model was trained on 8.4 trillion tokens before extended reasoning phases out to 1 million token trajectories.

Inkling 2 Small scored 31 on the Artificial Analysis Intelligence Index. On SWEBench Pro Hard, the model achieved 86.3 percent. On DeepSWE, the score is 72.1 percent. These numbers show this model is strong in programming and analysis tasks.

The company just closed a $5 billion funding round led by Accel at a $40 billion pre-money valuation. That is more than three times the roughly $12 billion valuation from last year's $2 billion seed round. Nvidia is separately in talks to invest about $2.5 billion.

MoE: Only 17 Billion Active out of 594 Billion

The figure that sets this model apart is the ratio between total and active: 594 billion total parameters, with only 17 billion switched on per token. A ratio of about 35 to 1 is the hallmark of Mixture-of-Experts, an architecture made of many small sub-networks called experts, where only a few are picked each time the model answers.

A coffee-shop analogy helps: imagine a kitchen with 35 cooks, but each order is handled by the single cook who knows that dish best. The staff totals 35 people, yet each customer is served by one. The customer pays for one cook's work, even though the kitchen employs 35.

This matters for inference cost: processing one token activates only 17 billion parameters, so the compute cost per token is far cheaper than a dense 594-billion model. That is the long-promised MoE advantage, and it is why the 594 billion headline sounds huge while the bill stays small.

The trade-off is memory: all 594 billion parameters still have to sit in GPU memory before the model can run, because any expert could be called at any time. MoE models are cheap to use but memory-hungry, a real limit for anyone who wants to run one at home.

A 292-Billion Vision Encoder and a Million-Token Trajectory

This model carries 292 billion parameters dedicated to vision, an encoder that turns images and video into input the language model can read. That is not a small number: it is roughly 17 times the size of its text path. This puts Inkling 2 Small in the serious multimodal class, not a text model with a picture feature bolted on.

The training length is also remarkable. The model was trained on 8.4 trillion tokens, then continued through extended reasoning phases out to 1-million-token trajectories. One million tokens is roughly enough to hold several short novels in a single uninterrupted reasoning session.

Note that this length describes the reasoning trajectory during training, not a claim that the model answers that long in every conversation. Every answer is still billed by its actual length, so long usage stays expensive even though the per-token price is cheap.

Tinker: A Closed Model on a Semi-Open Platform

The unusual release path of Inkling 2 Small is that it launched first through Tinker, Thinking Machines' own platform, before its weights open up. On Tinker, developers can call the model over an API without downloading it. Once external safety testing finishes, full Apache 2.0 weights are planned for release next week.

The sequencing makes sense from a security standpoint: a newly launched 594-billion-parameter model is a large attack surface. Testing it behind an API first gives the company time to watch how it behaves in real users' hands before the weights spread out of control.

The risk of this approach is just as clear: while the weights are unavailable, the open-weight community cannot run the model itself, replicate its results, or test its benchmark claims. For now, every number the public reads comes from the company and its evaluators, not from anyone's independent run.

What Apache 2.0 Means for Readers

The Apache 2.0 license is the most permissive of the popular licenses: anyone may use, modify, and distribute the model even commercially, with no licensing fees. The only requirements are the usual three: give attribution, show what you changed, and do not imply the owner's endorsement.

If the weights really land next week, developers could fine-tune Inkling 2 Small for their specific cases, run it on their own infrastructure, and be free of dependence on anyone's API. That is very different from "research only" licenses other large models often use.

Two cautions are worth keeping. First, the wording is "planned" and "after safety testing", which means the schedule can slip and the status stays uncertain until the weights actually exist in a repository. Second, no outside party can yet verify these launch numbers, because nobody can download and test the weights.

Deeper Dive: Thinking Machines' Position in the AI Model Market

Thinking Machines Lab was founded by Mira Murati after she left OpenAI. The company quickly became one of the most valuable AI startups in the world. With $5 billion in funding, they have resources to compete with large labs like OpenAI, Anthropic, and Google DeepMind.

Here is a comparison with other open-weight models:

Inkling 2 Small offers relatively small active parameters compared to total. This makes the model efficient in inference. But weights are not available yet, so the community cannot run it locally.

Implications for AI startups are also significant. With more open-weight models available, startups are not dependent on just one vendor. They can try various models to find the one that best suits their needs.

For developers, the release of open-weight models from Thinking Machines provides a new option. Local developers can wait for Apache 2.0 weights to try this model without high licensing costs. However, availability still depends on the weight release schedule. The compass has already shifted: the question is no longer whether large open-weight models will exist, but how quickly the weights can actually be downloaded and tested by the public.

Impact on AI Startups

The arrival of Thinking Machines Lab as a new player in the AI market gives impact to the AI startup ecosystem. First, more open-weight models available means local startups are not dependent on just one vendor. Second, increasingly tight competition between large AI labs can drive down model usage costs for startups. Third, Thinking Machines Lab's success in getting large funding shows investors are still optimistic about the AI sector.

However, startups still face several challenges. Computing infrastructure limitations remain a big barrier. GPU rental costs for training and inference are still expensive for many local startups. In addition, experienced AI talent is still relatively limited.

Therefore, the release of Inkling 2 Small is positive news for the AI ecosystem, but more support is needed from various parties to maximize its benefits.

Key Terms Glossary

• Thinking Machines Lab: AI company led by Mira Murati
• Mira Murati: Former OpenAI CTO
• Inkling 2 Small: 594 billion parameter MoE model
• Apache 2.0: Open source license allowing commercial use
• MoE: Mixture-of-Experts, AI model architecture
• SWEBench: Benchmark for testing AI programming capability
• DeepSWE: More difficult programming benchmark
• Tinker: Thinking Machines' API platform for running models in closed form