In brief
Thinking Machines Lab released Inkling on July 15—a 975-billion-parameter multimodal AI model trained from scratch, with full weights on Hugging Face under an Apache 2.0 license.
Inkling scores 74.1% on MCP Atlas—nearly 30 points above Nvidia's Nemotron 3 Ultra—making it the best-performing Western open-weights model on agentic tool use. Chinese models GLM 5.2 and Kimi K2.6 still lead on several key benchmarks.
Thinking Machines raised $2 billion at a $12 billion valuation in July 2025, then reportedly sought a $50 billion raise in November before those talks fell apart by January 2026.
Mira Murati left OpenAI in September 2024 to do her own thing. Almost two years later, that exploration shipped. Thinking Machines Lab, the company she founded, released Inkling—a multimodal AI model trained entirely from scratch, with every weight available for free download.When OpenAI's board fired Sam Altman in November 2023, Murati—then CTO—was named interim CEO. Altman was reinstated five days later, Murati returned to CTO, then left for good roughly 10 months after that. She founded Thinking Machines Lab in February 2025.The company then went quiet—and rich. It raised $2 billion at a $12 billion valuation in July 2025, led by Andreessen Horowitz with Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street alongside—one of the largest seed rounds in Silicon Valley history at the time.Reports in November 2025 had the company seeking a new round at a $50 billion valuation. Those talks collapsed by January 2026.What Inkling isInkling is a mixture-of-experts model—an architecture where only a portion of the network activates for any given input, keeping inference fast without sacrificing depth. It is a very big model: It has 975 billion total parameters (the internal settings that define how the model processes information), with 41 billion active per task, so forget about running it on your local machine.Being multimodal, this model accepts text, images, and audio, and supports a context window—the amount of text the model can reason over at once—of 1 million tokens, roughly 750,000 words. It was pretrained on 45 trillion tokens spanning text, images, audio, and video."Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today," Murati wrote on X . The fact that it’s trained from scratch means a lot, especially in the open-source community as it could bring a breath of fresh air to Western developers that are wary of China but need to use Asian models for their developments because the top AI companies in the Western world are mostly focused on shipping close-source models.Fine-tuning is the process of retraining an existing model on a specialized dataset to improve its performance on a specific task. Tinker is Thinking Machines' cloud platform built around that use case. The full weights are also on Hugging Face under an Apache 2.0 license, no restrictions.Inkling's clearest wins come in agentic tasks. On MCP Atlas—which measures how reliably an AI agent completes real-world tasks using Model Context Protocol, the open standard for connecting AI assistants to external tools and services, scored as percentage of tasks completed—Inkling posts 74.1%. That's nearly 30 points above Nvidia's Nemotron 3 Ultra, the main Western open-weights rival in the comparison.On SWE-Bench Verified—a test of whether an AI agent can autonomously fix real GitHub software bugs, scored as a percentage of issues resolved—Inkling scores 77.6%, also above Nemotron's 70.7%.Overall, Thinking Machines is selling this model as “well-rounded” and generalist. It means it doesn’t compromise quality in one specific set of tasks because its capabilities focus on something else (like models that are great at coding but suck at creative writing, for example).Source: Thinking MachinesThe Chinese models still have the edge on several fronts. Z.ai's GLM 5.2 scores 82.7% on Terminal Bench 2.1—a benchmark measuring autonomous AI coding agents in a real terminal environment, scored as percentage of tasks completed—against Inkling's 63.8%. Kimi K2.6 leads on Humanity's Last Exam, a test of PhD-level scientific reasoning.Thinking Machines acknowledges this. Inkling is not the strongest model available today, open or closed.What it is, is the most capable open-weights model built by a Western lab. Developers who—for legal, security, or compliance reasons—won't route workloads through models built in Beijing now have a real alternative to self-hosting Chinese models.Now, these developers have a model that (even though worse than the best Chinese models at almost everything) aligns better with their ideals, expectations and values. Subsequent finetunes can make this model excel at specific tasks, making those finetunes competitive in benchmarks versus Asian models.On FORTRESS Adversarial—which tests how consistently a model refuses genuinely harmful prompts without over-blocking legitimate ones, scored as a percentage correctly handled—Inkling scores 78.0%, the highest mark among all open-weights models in the comparison.Alongside Inkling, Thinking Machines previewed Inkling-Small: 276 billion total parameters, 12 billion active, already matching the larger model on most reasoning benchmarks. Its weights arrive once testing is complete, with no timeline given.Daily Debrief NewsletterStart every day with the top news stories right now, plus original features, a podcast, videos and more.
