Inkling 2 Small boosts AI task automation
Thinking Machines Labs has introduced Inkling 2 Small, a model designed to automate AI training and work tasks. The launch is accompanied by a $5 billion fundraising reported by Business Insider.
Thinking Machines Labs has unveiled Inkling 2 Small, a large language model designed to automate AI training and work tasks, while Business Insider simultaneously reports a $5 billion funding round led by Accel [to be verified]. Founded last year by Mira Murati, former CTO of OpenAI, the young company aims to prove it can already compete with the big players. The model is available at launch via the Tinker cloud service, with an open release of the weights announced for next week.
A model designed to do the dirty work for AI teams
Inkling 2 Small does more than just generate text. Thinking Machines says it trained the model for long and technical tasks, including "autoresearch" exercises that involve post-training networks, optimizing kernels — very low-level computation blocks — and conducting other research work on models. In the industry, this type of tool matters because it can reduce time spent on repetitive tasks while leaving fine judgments to human teams.
The company also claims to have tested the model on forecasting tasks. Specifically, Inkling 2 Small was used to fine-tune Qwen3.5-4B to improve predictions such as weather forecasts. According to Thinking Machines, the model's forecasting score doubled in 4.5 hours. This figure is provided by the publisher itself but gives an idea of the intended use: speeding up costly post-training cycles in terms of computation and time.
What the technical sheet reveals about Thinking Machines' strategy
The model was trained on a cluster of GB300 NVL72 servers, each integrating 72 Blackwell Ultra GPUs and half with CPUs. The initial pre-training reportedly took 68 hours, followed by a nine-day reinforcement learning workflow with several intermediate steps. These details matter because they indicate the level of infrastructure the company already mobilizes for a next-generation product.
Inkling 2 Small features 594 billion parameters in total, including 292 million for the visual encoder. About 17 billion parameters activate per query, allowing computational power to be adjusted as needed. In other words, the model can target either quality or efficiency. Not always both at the same time; the machine doesn’t have superpowers either.
Why the visual component changes the game
The visual encoder enables the model to analyze images and verify the quality of what it generates. Thinking Machines explains that a user can ask it to create an e-commerce store, then have it check both the code, any bugs, and the display of product pages. This capability directly interests business use cases because it brings content generation closer to quality validation.
The model has also been compared to Spark Open, Kimi K3, and other open-source models on several intellectual work benchmarks, where it achieves better scores according to the company. Thinking Machines also says it outperforms other references on several datasets dedicated to financial tasks. These comparisons should be taken with caution until full protocols are published, but they set the tone: the startup aims less to be spectacular than to establish itself on measurable professional uses.
Key points
- Late Thursday presentation, according to SiliconANGLE, on November 6, 2026.
- Business Insider mentions $5 billion raised.
- Inkling 2 Small has 594 billion parameters.
- The initial pre-training lasted 68 hours.
- Thinking Machines plans open-source weights next week.
- Mira Murati founded Thinking Machines last year.
By the numbers
- $5 billion — funding reported by Business Insider, 2026 [to be verified].
- 68 hours — duration of initial pre-training, according to Thinking Machines.
- 9 days — reinforcement learning workflow, according to Thinking Machines.
- 594 billion — total model parameters, according to Thinking Machines.
- 17 billion — parameters activated per query, according to Thinking Machines.