Introduction¶

The R&D pipeline for large models is extensive: pre-training, reinforcement learning, and post-training for kernels and downstream tasks often require significant manpower and dedicated scripts. Thinking Machines Labs Inc. (founded by former OpenAI Chief Technology Officer Mira Murati last year) released the new large model, Inkling 2 Small, around November 6, 2026. One of its positioning goals is to enable the model to participate in “model R&D”—such as long-range programming, self-developed post-training tasks, and optimization work on LLM components (including kernels). It was initially made available via the company’s Tinker cloud service and plans to release the weights under an open-source license within a week. On the same day, SiliconANGLE, citing Business Insider, reported that the company is advancing a financing round led by Accel, valued at approximately $5 billion with a valuation of about $40 billion; this part consists of media reports and is not an official company announcement.

What is Inkling 2 Small¶

Inkling 2 Small is a Large Language Model released by Thinking Machines aimed at “automating AI training-related tasks.” According to a SiliconANGLE report dated November 6, 2026, the company trained using GB300 NVL72 rack clusters: each system contains 72 Blackwell Ultra GPUs and a corresponding number of CPUs. The pre-training phase was completed in approximately 68 hours, followed by a reinforcement learning process lasting about nine days, which included several intermediate training steps.

In terms of parameter scale, the report gives a total of approximately 594 billion (written as 594 billion in the text), with approximately 17 billion active parameters per inference; the visual encoder accounts for about 292 million parameters. The model can adjust computing power investment based on the task, facilitating a trade-off between output quality and cost.

Training Data and Capabilities¶

Thinking Machines reinforced the programming capabilities of Inkling 2 Small through extensive long-range programming challenges. The tasks included high-difficulty “autoresearch” exercises: requiring the model to perform post-training on neural networks, optimize LLM components known as kernels, and execute other AI research work.

The company demonstrated its machine learning capabilities through a practical test: instructing the model to fine-tune the open-source model Qwen3.5-4B within 4.5 hours to enhance prediction capabilities such as weather forecasting. According to the company, this experiment improved the relevant prediction scores by approximately double. Additionally, reports indicate it performed better on three knowledge work benchmarks involving open-source models such as Spark Open and Kimi K3, and scored higher on several financial task datasets.

The model integrates a visual encoder capable of analyzing images and supports visual quality checks on graphics generated by the model itself. For example, if a user requests the construction of an e-commerce site, in addition to checking code, the model can also perform visual verification of product page rendering.

Developers familiar with the post-training process can customize Inkling 2 Small using their own data. The company states that compared to early versions, this model learns faster, which helps reduce the cost of running post-training.

Access and Open Source Plans¶

On the day of release, Inkling 2 Small was made accessible via Thinking Machines’ Tinker cloud service. The company plans to release the model weights under an open-source license one week later. Specific license terms and repository address are subject to official release at that time.

Funding Rumors and Industry Focus¶

Coinciding with the product release, SiliconANGLE, citing Business Insider, reported that Thinking Machines is raising approximately $5 billion, led by Accel, with an expected company valuation of about $40 billion. The aforementioned amounts and valuation have not been verified in the company’s official disclosures found in the retrieved documents, so it is advisable to mark them as “as reported” in writing and citation.

In terms of the product line, Inkling 2 Small brings the “agentic training stack” from research narrative to accessible models and upcoming open weights: long-range coding, post-training, and kernel optimization are incorporated into training objectives and public cases (such as the Qwen3.5-4B fine-tuning experiment). If the open-source weights are released as scheduled, the community will be able to directly reproduce or compare its performance on autoresearch tasks; the cloud-side Tinker service will handle early trials and commercial access. If the financing rumors are true, they reflect capital’s interest in betting on the “model-assisted model R&D” approach by the Murati team, but ultimately, this will depend on company announcements and product iterations.