Thinking Machines' Inkling-Small scores 80.2% on SWE-Bench Verified

Thinking Machines released Inkling-Small on July 30, a coding model that activates just 12 billion of its 276 billion total parameters on every token, according to the company.

That's the headline number: 80.2% on SWE-Bench Verified from a model running a small fraction of its parameters per token. Inkling-Small also posted 31.6% on Humanity's Last Exam and 82.2% on IFBench, per the release.

The model runs a 1-million-token context window and works natively across text, image, and audio. Reasoning effort is adjustable per request, according to Thinking Machines.

Weights are open on Hugging Face, and fine-tuning runs through Thinking Machines' own Tinker service. Output tokens cost $1.20 per million, per the company. The release doesn't list a price for input tokens.

Open weights and Tinker fine-tuning mean teams can test Inkling-Small against their own coding workloads before committing budget. Whether the $1.20 output price holds up against real usage patterns is the open question the release doesn't answer.