Thinking Machines Lab Releases Inkling, 975B-Parameter Open-Weights Multimodal AI Model

Jul 15, 2026

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released Inkling, a 975-billion-parameter open-weights Mixture-of-Experts model under an Apache 2.0 license. Designed for customization rather than raw leaderboard dominance, Inkling natively processes text, images, and audio. It introduces a controllable thinking effort mechanism allowing developers to scale reasoning budgets. While trailing top proprietary models, Inkling delivers competitive agentic performance and strong censorship resistance.

Inkling model architecture release

▪Thinking Machines Lab released Inkling, an open-weights Mixture-of-Experts transformer model with 975 billion total parameters and 41 billion active parameters, on July 15, 2026.
▪Inkling supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens of text, images, audio, and video.
▪Inkling is released under the permissive Apache 2.0 open-source license, allowing developers to download, modify, and commercialize the model weights royalty-free.
▪Thinking Machines Lab is previewing Inkling-Small, a lighter-weight Mixture-of-Experts model with 276 billion total parameters and 12 billion active parameters.

Multimodal reasoning capabilities

▪Inkling natively processes text, images, and audio using an encoder-free architecture that projects all modalities into a shared hidden space.
▪Inkling scored 77.2% on the MMAU audio benchmark and 91.4% on VoiceBench, compared to Gemini 3.1 Pro's 82.5% on MMAU and 94.4% on VoiceBench.
▪Inkling inputs audio signals as discrete dMel spectrograms and encodes images as 40x40 pixel patches using a four-layer hierarchical multi-layer perceptron.

Agentic coding tool use

▪Inkling refined an online multiplayer snake game through 40 iterations of feedback from GPT Codex serving as an external reviewer.
▪Inkling scored 77.6% on SWE-bench Verified, outperforming Nvidia Nemotron 3's score of 71.9% but trailing Claude Fable 5's score of 95.0%.
▪Inkling ranks among the strongest open-weights models on Design Arena's Agentic Web Dev leaderboard with a human evaluation score of 1257.

Controllable thinking effort

▪During reinforcement learning training over 30 million rollouts, Inkling naturally compressed its internal reasoning steps to reduce latency, a phenomenon called chain of thought condensation.
▪Inkling features a controllable thinking effort mechanism that allows developers to programmatically adjust the model's reasoning budget from 0.2 to 0.99.

Fine-tuning customization platform

▪To demonstrate customization, Inkling autonomously wrote, executed, and evaluated its own fine-tuning job using the Tinker platform.
▪Thinking Machines Lab partnered with platforms including Together, Fireworks, Modal, Databricks, and Baseten to provide API deployment for Inkling.
▪Inkling is available for fine-tuning on Thinking Machines Lab's Tinker platform, which includes an Inkling Playground chat interface in the console.

Epistemic calibration training

▪Inkling scored 98.6% on the StrongREJECT benchmark and achieved a 78.0% refusal rate on adversarial queries on the FORTRESS benchmark.
▪Inkling was evaluated on Cognition's Propaganda and Censorship Eval, where it demonstrated patterns of censorship non-compliance by answering directly on sensitive topics.
▪Inkling's instruction following was trained using a rubric grader for recall and a claims grader that performs agentic web searches to verify factual claims.
▪Inkling was trained for calibration using reinforcement learning against proper scoring rules on a large corpus of resolved real-world questions.