Thinking Machines releases Inkling, 975B-parameter open-weight model

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released its first AI model Wednesday: Inkling, a 975-billion-parameter mixture-of-experts system that developers can download and modify directly. Unlike flagship models from OpenAI, Anthropic, or Google, Inkling is open-weight.

Technical specifications

Inkling uses a mixture-of-experts architecture with 975 billion total parameters but activates only about 41 billion for any given task. The model was trained on 45 trillion tokens spanning text, image, audio, and video data. It reasons natively across all modalities, according to the company.

The model includes adjustable "thinking effort" — users can dial up reasoning depth or prioritize speed. On coding benchmarks, Thinking Machines claims Inkling uses one-third as many tokens as Nvidia's Nemotron 3 Ultra to achieve equivalent performance.

Positioning and strategy

Thinking Machines explicitly states in its briefing materials that Inkling is "not the strongest model available today, closed or open." Instead, the company is positioning it as a customization starting point through Tinker, its model fine-tuning platform.

The core thesis: centralized AI labs sell one-size-fits-all models, while enterprises that customize their own models extract more value. Microsoft CEO Satya Nadella made a similar argument Sunday, noting that enterprises using proprietary models "pay twice" — subscription costs plus business knowledge embedded in prompts.

Training and costs

Thinking Machines pretrained Inkling from scratch but used outputs from other open-weight models, including Moonshot AI's Kimi K2.5, to generate some early post-training data before large-scale reinforcement learning. The company says its next model will use fully self-contained post-training.

The model was trained entirely on Nvidia GB300 NVL72 systems through a strategic partnership announced in March for one gigawatt of Vera Rubin computing capacity. Nvidia made a "significant investment" in Thinking Machines at that time. Pricing for model access has not been disclosed.

A reported $50 billion fundraising round was coming together in November 2025 but had stalled by January 2026, according to multiple outlets. The company has declined to discuss funding since.

Evidence for the approach

In a joint project with Bridgewater Associates (not an investor), researchers fine-tuned an open-source model on Bridgewater's financial expertise. The result scored 84.7% on financial reasoning tests, outperforming top proprietary models while costing roughly one-fourteenth as much to run. Those results come from the companies' own evaluation, not an independent benchmark.

Company status

Thinking Machines now employs approximately 200 people, up from earlier 2026 levels after departures including two co-founders who left for OpenAI in January. The company took roughly nine months from formation to release and revenue, compared to five years for OpenAI and three for Anthropic, according to the company.

What this means

Inkling tests whether enterprises will adopt open-weight models they can customize over proprietary APIs. The revenue model depends on Tinker's fine-tuning platform rather than metered API access, since anyone can download and run the weights independently. If Thinking Machines is correct, frontier labs' business models face pressure from organizations building in-house AI capabilities — but the company must prove that customization value exceeds the convenience of ChatGPT or Claude.