US-based Thinking Machines Releases Inkling 2 Small, a 594-Billion-Parameter AI Model

en.Wedoany.com Reported - On November 5, the US artificial intelligence company Thinking Machines Lab officially launched its second-generation AI model, Inkling 2 Small. The model uses a mixture-of-experts (MoE) architecture, with 594 billion total parameters and 17 billion activated parameters per inference, and supports text and image inputs as well as adjustable reasoning intensity. The new model is already available for inference and fine-tuning through the company's Tinker platform, and the full model weights are planned to be released under the Apache 2.0 license after external safety testing is completed.

Inkling 2 Small uses a 48-layer Transformer architecture, activating 8 of 384 expert networks per token and combining local and global attention mechanisms. The model supports a context window of up to 1 million tokens and supports the MXFP8 and NVFP4 numerical formats, and can be used for software development, agentic tool calling, knowledge retrieval, and complex reasoning tasks.

According to technical data published by Thinking Machines, the first stage of pretraining used 8.4 trillion tokens, ran for 68 hours on NVIDIA GB300 NVL72 computing systems, and used native NVFP4 format training. The second stage used 3.8 trillion tokens of high-quality data, and subsequent stages continued to introduce long-context reasoning data and agent operation trajectories, with the third stage using approximately 320 billion tokens of data and the fourth stage using approximately 34 billion tokens of data.

In the reinforcement learning stage, Thinking Machines conducted 9 days of joint training using 2,048 GPUs, completing a cumulative 1,075 training steps and more than 17 million task rollouts, processing more than 3 trillion tokens, with a compute volume exceeding 4×10²³ FLOPs. The training tasks covered code development, terminal operations, document generation, scientific reasoning, visual understanding, and tool calling.

Performance evaluation data shows that Inkling 2 Small scored 30.6 points on the Artificial Analysis Intelligence Index v5.0 evaluation, which the company's press release rounded to 31 points. On the AA-Briefcase v1.1 knowledge work test, the model scored 1,527 points; on the JobBench test, it achieved a result of 70.5%. The above evaluation results come from the test tables published by the company.

Inkling 2 Small is now available for developers to experience through the Tinker Playground and to conduct model inference and fine-tuning through Tinker. Thinking Machines plans to release the full model weights and model checkpoints the following week after external safety testing and red-team evaluation are completed. The currently released version provides services with closed weights.