Thinking Machines Launch Inkling 2 Small and fold LLM open source performance in 4.5 hours

Thinking Machines presented Inkling 2 Small, an AI model focused on business tasks automation and fine-tuning open source LLMs. According to the company, the model doubled the forecast score of a Qwen3.5-4B in just 4.5 hours — a result that the company uses to defend that compact models, well-adjusted, provide work of knowledge without relying on closed giants. The announcement takes place amid the search for the startup for a round of reported funding at $5 billion, according to SiliconANGLE.

What is Inkling 2 Small and why does it matter?

The position is specific: instead of competing for the position of the larger generalist model, the Inkling 2 Small targeting two practical fronts — automate business flows and serve as a basis for fine-tuning existing open models. The name indicates the bet on efficiency: the proposal is to extract performance from concrete tasks with a smaller model, adjusted to the company domain, rather than resort to a giant model for any problem. The source does not detail architecture, number of parameters, use license or access price.

The Inkling 2 numbers Small: Double performance in 4.5 hours

The central data released is the jump obtained over a Qwen3.5-4B: the forecast score doubled after 4.5 hours of adjustment. The company also claims that the model surpassed open competitors in knowledge work benchmarks — the category that includes reading, classification, forecasting and production of structured content. It is worth recording what was not informed: the specific benchmarks used, the absolute scores, the hardware used in training and the cost of each hour of adjustment. Without these data, comparison with rivals cannot yet be independently reproduced.

Fine adjustment of open LLMs: the new front of the AI dispute

The movement follows a major change in the market: companies started to prefer to adapt a model to their own domain rather than rely only on prompts in closed APIs. Models with open weights, such as Mistral Large 4 (Le Chonk), expanded this margin of manoeuvre and created demand for tools that make adjustment fast and cheap. It is exactly this space that Thinking Machines tries to occupy — and that also explains the interest of investors in a round of this size.

Impact of Inkling 2 Small for developers in Brazil

For small teams, the appeal is the short cycle: 4.5 hours of adjustment fit into a work window and do not require a dedicated cluster for weeks, which reduces the entry barrier for teams that do not have large infrastructure. In sectors such as legal, financial and care, the possibility of specializing an open model tends to weigh more than the adoption of a generic assistant. There is also the dimension of compliance: maintaining the adjustment and inference in its own environment facilitates processing personal data under LGPD, a point that gains relevance as corporate adoption advances.

Inkling 2 Small: gaps and unanswered questions

There is lack of information that decides to adopt in practice: price, license terms, limits on commercial use, availability outside the United States and cases of customers with verifiable metrics. Nor is there official confirmation about the $5 billion round — the amount was reported by the press, not announced by the company. And, like every benchmark released by the developer herself, Inkling 2 numbers Small requests verification by third parties before becoming a purchasing decision.

Frequently Asked Questions about Inkling 2 Small

What is Inkling 2 Small?

It is a model of Thinking Machines focused on automation of business tasks and fine tuning of open source LLMs, focusing on knowledge work. The company claims that it doubled the forecast score of a Qwen3.5-4B in 4.5 hours.

How much does it cost and how to access the model?

This information has not been disclosed. The source does not bring price, license or availability details for developers.

Is the $5 billion round confirmed?

Nope. The value was reported by the press and the announcement of the model occurs while the company seeks this capture, without official confirmation disclosed.