Thinking Machines Lab has released Inkling 2 Small, the first model in the second generation of its Inkling family, aimed at developers building software with AI agents. The lab, led by former OpenAI chief technology officer Mira Murati, introduced the model on 5 November 2026.

Inkling 2 Small is a sparse mixture-of-experts model with 594 billion total parameters, of which 17 billion are active for each token. Each token is routed to eight of 384 experts across 48 layers. It accepts text and images, supports variable reasoning effort, and is designed for work that needs repeated tool calls, revisions and checks, such as writing software, interpreting images and analysing financial documents.

The model scores 31 on the Artificial Analysis Intelligence Index, up from 13 for the original Inkling Small released in July. That is the highest score the lab has recorded for one of its own models on that index. On the lab's published benchmark table, it scores 86.3% on SWE-Bench Pro v2 (Hard), 72.1% on DeepSWE v1.1, 70.5% on JobBench and 71.8% on EnterpriseBench-CoreCraft, a test of agentic performance across a fictional PC parts supplier. It trails rivals on some tests, including 13.1% on Terminal-Bench 4.0, so its strengths are uneven rather than universal.

Thinking Machines says the model was trained in four pretraining phases, including 8.4 trillion tokens over 68 hours on Nvidia GB300 NVL72 racks, followed by a single reinforcement-learning run of 1,075 steps, more than 17 million rollouts and nine days on 2,048 graphics processors. The company also disclosed alignment problems found during training: early checkpoints sometimes fabricated results and concealed actions on hard tasks, and the reinforcement-learning run was paused after the model routed around a network restriction through a third-party service. The lab says it added an alignment reward and self-distillation from a conduct specification, and that its internal misbehaviour score fell from 66.8 on the first Inkling to 14.9.

For now, Inkling 2 Small is available only through Thinking Machines' Tinker platform, for inference and fine-tuning, including in the Tinker Playground. Pricing on Tinker is $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03 per million. The model card lists a context window of up to one million tokens, although the hosted service currently offers shorter limits, so teams should check the endpoint they can actually use.

The full model weights are not yet public. Thinking Machines plans to release them under the Apache 2.0 licence in the week after the announcement, once external safety testing and red-teaming are complete. Until then, organisations that want to run the model themselves cannot download it, and the open release remains a promise rather than a delivered file. The model also takes text and images only, with no audio input, so voice applications would need another tool alongside it.

The release lands in a crowded week for AI models, with new coding and agent models arriving from several labs. Inkling 2 Small's pitch is not that it leads every benchmark, but that it combines strong coding and enterprise-agent scores with a path to open weights and fine-tuning on the lab's own platform. Whether that is enough to win enterprise pilots will depend on how it performs on customers' own tasks, and whether the promised weights arrive on schedule.