Mira Murati Releases 975B Open-Weights Frontier AI Model
Former OpenAI CTO Mira Murati released Inkling on July 15, 2026: a 975 billion parameter open-weights model under Apache 2.0, the largest from a Western lab, positioned as customizable over raw capability.
The person who built OpenAI's closed system just shipped the exit ramp.
Key takeaways
- Thinking Machines Lab released Inkling on July 15, 2026: a 975 billion parameter mixture-of-experts model under an Apache 2.0 license, free to download, fine-tune, and redeploy.
- Inkling is explicitly not the most capable model available. The bet is customizability and cost asymmetry over benchmark leaderboard position.
- The release fills a gap left by Meta's retreat from open Llama development and directly challenges the closed-model subscription model that OpenAI and Anthropic have built their businesses on.
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati in February 2025, released Inkling on July 15, 2026. It is a 975 billion total parameter mixture-of-experts transformer, 41 billion parameters active per forward pass, pretrained on 45 trillion tokens of text, images, audio, and video. Full weights are publicly available under an Apache 2.0 license. The lab built Inkling from scratch, in about nine months, with roughly 200 employees.
The company's own statement sets the bar honestly: "It is not the most performant model available today, closed or open." The explicit pitch is something else entirely.
What Inkling Actually Is
The model supports a 1 million token context window and was trained on Nvidia GB300 NVL72 systems, part of a partnership with Nvidia for a gigawatt of Vera Rubin computing capacity announced on March 10, 2026 (Nvidia blog, Thinking Machines announcement). A quantized NVFP4 version is available that cuts GPU memory requirements in half. Inference is live on Together AI, Fireworks, Modal, Databricks, and Baseten. An Inkling-Small variant at 12 billion active parameters is in preview simultaneously.
Thinking Machines does not monetize Inkling directly. Revenue comes through Tinker, its fine-tuning and customization platform. Inkling can write its own fine-tuning scripts to refine its behavior, demonstrated live on Tinker. Bridgewater Associates is already a Tinker customer.
The model's pretraining used some data generated by other open-weight models, including Moonshot AI's Kimi K2.5, before large-scale reinforcement learning took over. That partial distillation is worth noting, though it does not undermine the core claim: Inkling was built from scratch by a Western team, with full weights released publicly, under the most permissive license available.
The Closed-Model Cartel Has a Real Problem Now
The power concentration angle is the story the benchmark coverage misses. OpenAI and Anthropic have constructed a subscription tollbooth on frontier intelligence. Every enterprise running prompts through their APIs hands over not just money but proprietary data embedded in those interactions, which feeds back into the closed model. That's the loop. Open weights breaks it. Your weights, your data, your infrastructure. Self-custody applied to AI.
The Western open-weights vacuum has been a real and growing problem. Meta retreated from Llama after a lackluster Llama 4 release. The gap it left has been filled largely by Chinese models: DeepSeek V4, GLM 5.2, Kimi K2.6. Inkling, at 975 billion parameters, is the leading open-weights release from a U.S. lab by score on the Artificial Analysis Intelligence Index (Artificial Analysis), scoring above Nvidia's Nemotron 3 Ultra, the previous benchmark. For any builder who cares about not defaulting to CCP-adjacent infrastructure, that distinction matters. U.S. export controls have already handed Chinese AI significant wins by restricting Western alternatives. The export control dynamic is accelerating the urgency of credible Western open alternatives.
Murati was at OpenAI in 2019 when the lab first withheld the full GPT-2 release over misuse fears. That was the beginning of the closed-model playbook she watched get built from the inside over the following five years. She left in September 2024, founded Thinking Machines in February 2025, and built a competing open system in about nine months. That's a revealed preference from someone with more inside knowledge of the closed-model regime than almost anyone alive.
The Tinker revenue model rhymes with open-source Bitcoin infrastructure. The protocol is free. The services and tools on top are where value accrues. Give the weights away, monetize the customization toolchain, and let builders own their stack. Thinking Machines is betting that customizability and cost asymmetry beat raw benchmark position for the majority of real enterprise use cases. That is a falsifiable thesis, not a press release.
The thesis fails if, within 12 to 18 months, developer adoption of Inkling stays negligible while OpenAI and Anthropic compound their distribution flywheels, or if the fine-tuning market proves too commoditized for Tinker to hold a defensible position. It also fails if enterprise buyers simply don't defect from closed subscriptions regardless of the customization argument. Those are the metrics to watch.
What Comes Next
Inkling is Thinking Machines' first public product. The lab has operated at a reported $12 billion valuation before any public release, described across multiple secondary sources as one of the largest seed rounds on record. Whether the Tinker platform converts that valuation into durable revenue, and whether the open-weights bet attracts the developer ecosystem needed to compete with closed incumbents on distribution, are the questions the next 12 months will answer. Watch Tinker customer growth, fine-tune adoption on Hugging Face, and whether enterprise buyers who currently pay OpenAI or Anthropic subscriptions start defecting to self-hosted Inkling deployments.
Sources
Frequently Asked Questions
Open-weights means the trained model parameters are publicly downloadable and modifiable. The training data and source code used to build the model are not necessarily released. Developers can run, customize, and deploy the model without access to the full build pipeline. Apache 2.0 is the most permissive license available: free to fine-tune and redeploy commercially.
At native precision, Inkling requires over 2 terabytes of GPU memory, roughly 16 Nvidia H200 accelerators. The quantized NVFP4 version cuts that requirement in half. Most individuals and small teams will access it through inference APIs on Together AI, Fireworks, Databricks, Baseten, or Modal rather than running it locally.
Murati was present at OpenAI through its full transformation from a research lab that withheld GPT-2 into a closed, for-profit model shop. Her choice to release Inkling with full open weights under Apache 2.0 is a direct repudiation of that path, backed by firsthand knowledge of why the closed model was built the way it was. The defection signal carries more weight coming from someone who built the system she's now routing around.
