Thinking Machines Launches Inkling 2 Small as Startup Seeks $5B Round

In total, the model encompasses 594 billion parameters. However, it operates on a sparse activation pattern where approximately 17 billion parameters are activated when processing an incoming prompt. Included within the overall architecture is a 292-million-parameter vision encoder. This vision component gives Inkling 2 Small the ability to parse imagery and visually evaluate the outputs it produces. For instance, when tasked with creating an e-commerce platform, the model can inspect its own code for software bugs while also visually reviewing generated product pages to confirm that elements render properly.

Automated machine learning and benchmark results

To build the model's technical problem-solving capabilities, Thinking Machines exposed Inkling 2 Small to hundreds of thousands of long-horizon programming problems. Among these were sophisticated "autoresearch" challenges requiring the system to optimize neural network kernels, post-train other models, and execute routine artificial intelligence research duties.

To evaluate its practical utility in machine learning workflows, the company directed Inkling 2 Small to fine-tune Qwen3.5-4B, an open-source model. The specific target was improving the model's accuracy on predictive tasks such as weather forecasting. According to Thinking Machines, Inkling 2 Small succeeded in doubling Qwen3.5-4B's forecasting benchmark score within 4.5 hours.

Beyond specialized machine learning engineering, the model was evaluated against traditional business and knowledge work applications. Thinking Machines reported that Inkling 2 Small outperformed competing open-source offerings—including Spark Open and Kimi K3—across three standard knowledge work benchmarks. The model also posted leading scores on a series of financial task datasets.

What it means for developers

For software engineers and machine learning practitioners, Inkling 2 Small introduces flexible operational controls and new automation possibilities. The model is built to let developers dynamically adjust the computational budget allocated to a given prompt. This allows teams to balance output accuracy against inference costs, tailoring resource usage to high-priority production tasks or higher-volume, budget-conscious operations.

Furthermore, Thinking Machines highlighted that Inkling 2 Small acquires domain-specific knowledge faster than its earlier architectures. This characteristic lowers the computational expense associated with customizing or post-training the model on proprietary datasets. Developers can train the system on internal data while spending fewer compute hours than prior iterations required.

As engineering teams explore newer architectures like Inkling alongside existing production systems, monitoring cost and inference quality remains critical. Developers can test and deploy top AI models cheaply through one API at https://apixoai.online to benchmark multiple leading options against their specific operational needs.

Inkling 2 Small is currently available through Thinking Machines' proprietary Tinker cloud service. The company has announced plans to release the model's weights under an open-source license next week, allowing the broader engineering community to inspect and run the system locally.

Source: Thinking Machines debuts Inkling 2 Small amid reported $5B fundraising push — SiliconANGLE AI. Written by the Apixo team from that report.