Thinking Machines released Inkling-Small, an open-weight 276-billion-parameter model that matches its larger Inkling system on reasoning and coding benchmarks while activating roughly a quarter of the compute per token. The release gives developers and enterprises a cheaper option for near-frontier multimodal reasoning without the inference bill that comes with running a full-scale flagship. Thinking Machines, the AI lab founded by former OpenAI chief technology officer Mira Murati, built the smaller model to train faster and serve cheaper than Inkling itself.
Inkling-Small is a Mixture-of-Experts (MoE) transformer, an architecture that splits the network into specialized sub-networks called experts and routes each token to only a handful of them instead of computing through the entire model. Of its 276 billion total parameters, only 12 billion are active parameters, the portion the model actually runs for any single token; Inkling, for comparison, carries 975 billion total parameters with 41 billion active. That gap is the efficiency story: Inkling-Small activates less than a third of Inkling’s per-token compute while landing in the same performance range.
Thinking Machines trained Inkling-Small on Nvidia’s GB300 NVL72 systems, starting the run after Inkling’s own training finished so the team could apply lessons learned to the pre-training data mix and recipe. The company post-trained an earlier checkpoint using on-policy distillation, with Inkling itself serving as the teacher model, then scaled agentic coding reinforcement learning for two more weeks. Thinking Machines says that process pushed Inkling-Small past Inkling on reasoning and agentic coding scores, though Inkling retains an edge on knowledge coverage and factuality.
On the company’s own benchmarks, Inkling-Small scored 31.6% on Humanity’s Last Exam against Inkling’s 29.7%, cleared 80.2% on SWE-Bench Verified, and hit 64.7% on Terminal-Bench 2.1, a measure of agentic tool use. These figures come from Thinking Machines’ own evaluations, in part citing outside benchmark trackers Artificial Analysis, Scale AI, and the ARC Prize Foundation, not from an independent third-party audit. The company’s own data also shows where the smaller model falls short: Inkling-Small scored 20.6% on SimpleQA Verified against Inkling’s 43.9%, meaning it fabricates or forgets more raw facts even as it reasons and codes competitively.
Inkling-Small keeps Inkling’s multimodal design, encoding audio as spectrograms and images as small pixel patches so both feed into the same transformer alongside text, and it supports variable thinking effort, letting a developer dial reasoning depth up or down to trade cost against accuracy. Its context window extends to 1 million tokens, matching Inkling’s, supporting long-document or long-audio tasks that shorter-context models cannot hold in a single pass.
Thinking Machines is releasing Inkling-Small’s full weights and opening the model up so developers can fine-tune it directly on Tinker, its model-training platform, with chat access through Tinker Playground across text, image, and audio. The launch lands two days after AI Insiders reported that Thinking Machines co-founder Lilian Weng left the company citing health reasons and subsequently joined OpenAI. Nothing in the Inkling-Small release ties the two events together.
Teams currently paying for frontier-scale inference to get reasoning and coding quality they could get from a smaller model should run Inkling-Small’s benchmark set against whatever they deploy today. Any team weighing the switch should budget separately for the fact-accuracy gap the company’s own SimpleQA numbers reveal, since a model tuned for agentic reasoning at lower compute is not automatically one tuned to avoid factual error.
Thinking Machines announced Inkling-Small on its official site in late July 2026.
