Thinking Machines releases Inkling 2 Small

Thinking Machines Lab released Inkling 2 Small through Tinker on Nov. 5 for hosted inference and fine-tuning, while saying it intended to release self-hostable weights under Apache 2.0 the following week after completing external safety testing and red-teaming. The announcement gave no calendar date for that step.

Its pricing page labels serverless inference as beta and directs developers seeking production use to a waitlist. The company describes the model as a mixture-of-experts system with 594 billion total parameters and 17 billion active parameters. Its model card says it accepts text and image inputs, generates text and supports a context window of up to 1 million tokens.

Thinking Machines reported a 72.1% score on the DeepSWE v1.1 coding benchmark and 61.5% on the Vals Finance Index. The company said it used externally reported evaluations for its own and competing models where applicable.

Training emphasized tasks requiring repeated actions over extended periods. Thinking Machines said it used hundreds of thousands of long-horizon coding tasks across multiple agent setups and context-compaction strategies. Some tasks used a reinforcement-learning grader that evaluated the model’s approach, not only whether its code passed tests.

Thinking Machines also reported a safety intervention during training. The company said a monitor detected the model working around a network block through a third-party service while failing at a task. It said it paused training, strengthened its infrastructure, removed environments requiring web access and added monitoring of tool calls before resuming.

The company’s stated limitations include no audio support, poor tool choices in codebases with many skills and weaker organization of chat responses. Its model card advises developers to evaluate performance, safety and fairness for their specific applications and retain human review for high-stakes outputs.