Mira Murati’s Thinking Machines Releases Its First Open Weights Model, Inkling

Thinking Machines Lab, the AI startup founded in early 2025 by former OpenAI chief technology officer Mira Murati, released its first model on Wednesday: an open weights system called Inkling. Unlike the flagship models sold by OpenAI, Anthropic, and Google, Inkling’s full weights are public, so any developer or company can download, inspect, and modify the model directly. “Inkling is not the strongest overall model available today, open or closed,” Thinking Machines wrote in its own announcement, framing the 975 billion parameter system instead as a flexible base for other organizations to build on.

Inside Inkling’s Architecture

Inkling is a Mixture-of-Experts (MoE) transformer with 975 billion total parameters, of which about 41 billion activate for any given request, according to Thinking Machines’ announcement. That design, now common among very large models, keeps inference faster and cheaper than activating every parameter on every token: the company says each token routes through six of 256 routed experts plus two shared experts. The Register reports that Inkling’s mixture-of-experts approach was inspired by DeepSeek-V3, though Thinking Machines trained the model from scratch on Nvidia’s GB300 NVL72 systems, using 45 trillion tokens of text, images, audio, and video as pretraining data.

The model supports a context window of up to 1 million tokens, large enough to hold sizable codebases or document sets in a single session. Thinking Machines says Inkling reasons natively across text, images, and audio, though TechCrunch’s report on the release notes that for now its outputs are limited to text, including code, styled artifacts, and structured data.

Running the model at its native 16-bit precision takes more than two terabytes of GPU memory: roughly eight of Nvidia’s B300 accelerators, or sixteen H200s, according to The Register’s analysis. Thinking Machines also published an NVFP4-quantized version that fits on about half that hardware, along with a preview of a smaller sibling, Inkling-Small, a 276 billion parameter MoE model with 12 billion active parameters aimed at lower latency and cost. Inkling is released under the Apache 2.0 license, which lets anyone fine-tune, redistribute, or commercialize it without paying Thinking Machines a fee.

How It Stacks Up

Thinking Machines is pitching Inkling on well-rounded performance rather than raw benchmark leadership. On Terminal Bench 2.1, the company says Inkling matches the coding performance of Nvidia’s Nemotron 3 Ultra, previously the largest American open weights model at 550 billion parameters, while using roughly a third as many “thinking tokens” to get there. Those thinking tokens come from Inkling’s reasoning process: like most current large language models, it uses reinforcement learning to work through a chain of thought before answering, and users can dial that thinking effort up or down to trade speed for accuracy. Thinking Machines also designed the model to flag uncertainty rather than guess when it lacks a confident answer, a behavior the company discusses under what it calls the model’s “epistemics.”

The Register places Inkling’s scale alongside large Chinese open weights models such as DeepSeek V4, GLM 5.2, and Kimi K2.6, while noting that Thinking Machines’ own benchmark charts still show it trailing closed, proprietary models from Anthropic and OpenAI. Thinking Machines has also disclosed that although Inkling was pretrained from scratch, the company used other open-weight models, including Moonshot AI’s Kimi K2.5, to help generate some early post-training data before large-scale reinforcement learning, TechCrunch reported. As with any vendor-reported benchmark, these comparisons come from Thinking Machines’ own testing rather than independent evaluation.

Tinker and the Bridgewater Case Study

Thinking Machines is marketing Inkling less as a finished product and more as a starting point for other organizations to customize through Tinker, its model fine-tuning platform. That is a different approach than OpenAI, Anthropic, and Google have taken with ChatGPT, Claude, and Gemini, which compete first as general purpose chatbots. In a blog post published the week before Inkling’s launch, Thinking Machines argued that AI trained centrally by one company and then frozen in place underperforms models that organizations can shape with their own domain expertise, according to TechCrunch.

The company points to a project with Bridgewater Associates, the world’s largest hedge fund and not a Thinking Machines investor, as an example of that bet paying off. Researchers from both companies fine-tuned an existing open-source model further on Bridgewater’s financial expertise; the resulting model reportedly scored 84.7% on financial reasoning tests, ahead of top proprietary models, at roughly one-fourteenth the cost to run. TechCrunch reported that these figures came from the two companies’ own evaluation and have not been independently verified.

To show what self-customization looks like in practice, Thinking Machines had Inkling fine-tune itself as part of the announcement: the model wrote its own training objective and dataset, ran the job on Tinker, evaluated the result, and switched itself over to the updated weights. Thinking Machines also says it reached a market-and-revenue milestone in about nine months, versus roughly five years for OpenAI and three years for Anthropic to do the same, according to TechCrunch. The company has said comparatively little about its own funding since a reported $50 billion round that reportedly came together in November 2025 stalled by January 2026, and it has not detailed how it plans to cover the cost of training and serving frontier-scale models going forward.

Availability

Inkling is available now for chat and fine-tuning on Thinking Machines’ Tinker platform, including a new Inkling Playground interface in the Tinker console. Developers who want to run it themselves can download the full weights from Hugging Face, with day-one support from inference engines including vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp. Thinking Machines says it is also working with third-party hosts, including Together AI, Fireworks, Modal, Databricks, and Baseten, to bring Inkling to their platforms.

Sources

• The Register: Former OpenAI CTO does what Altman won’t: releases a frontier AI model that’s actually open
• Thinking Machines Lab: Inkling: Our open-weights model (official announcement)
• TechCrunch: Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling

Featured image: an Nvidia DGX GB200 rack with 72 B200 GPUs and 36 Grace CPUs, the same family of Nvidia liquid-cooled, rack-scale AI hardware as the GB300 NVL72 systems Thinking Machines used to train Inkling. Photo by Pokiiri, CC BY-SA 4.0, via Wikimedia Commons; cropped, resized, and converted to WebP.