# Thinking Machines releases Inkling 2 Small model

published: 2026-11-06T22:05:32.842Z
topic: infrastructure
source: https://siliconangle.com/2026/11/06/thinking-machines-debuts-inkling-2-small-amid-reported-5b-fundraising-push/
url: https://aiagentnews.io/stories/thinking-machines-releases-inkling-2-small-model

Thinking Machines Labs released Inkling 2 Small, a large language model it says can automate AI training tasks, and it is available now through the company's Tinker cloud service. The company also plans to release its weights under an open-source license next week.

Thinking Machines Labs released Inkling 2 Small late Thursday. The company trained it on GB300 NVL72 racks, completing pre-training in 68 hours and then a nine-day reinforcement learning workflow. Training also included hundreds of thousands of long-horizon coding challenges, including complex "autoresearch" tasks such as post-training neural networks and optimizing kernels.

In a test of its machine learning skills, the company had the model fine-tune Qwen3.5-4B to improve its forecasting, and the company says the forecasting score doubled in 4.5 hours. Thinking Machines also says Inkling 2 Small beat several open-source models, including Spark Open and Kimi K3, on three knowledge work benchmarks. A vision encoder, accounting for 292 million of the model's 594 billion parameters, lets it check how the web pages it generates render.

The model activates about 17 billion parameters per prompt and lets developers adjust how much compute each task uses, trading cost against output quality. Separately, Business Insider reported that Thinking Machines is raising a $5 billion round led by Accel at an expected $40 billion valuation.

## Why it matters if you build agents

Developers can use Inkling 2 Small through Tinker now and post-train it on their own data. Its adjustable compute setting lets agent builders balance cost and quality task by task.

The weights are not yet public, so check the open-source license once it ships before building production pipelines around the model.