Ex-OpenAI CTO Mira Murati's Thinking Machine Releases First Model Inkling

Thinking Machine, founded by former OpenAI CTO Mira Murati, has released its first model, Inkling, a 975B-parameter MoE architecture with 41B active parameters, a 1M context window, trained on 45T tokens, and natively multimodal (text, image, audio). The model is open-source, with post-training data distilled from models like Kimi 2.5. A smaller preview version, Inkling Small with 12B active parameters, is also available.

SynthePulse Insight · AI deep reading

Inkling Launch: Mira Murati's Open-Source Gamble and Distillation Controversy

Version 2 · 2 sources

Sep 2, 2026, 3:15 AM

Thinking Machine, founded by former OpenAI CTO, releases its first model Inkling, entering the open-source frontier with nearly a trillion parameters and 1M context, but its distillation data sources and performance positioning draw attention.

Inkling has 975B total parameters, 41B active parameters, uses MoE architecture, and has a context window of 1M tokens.
The model was trained on 45T tokens of data and natively supports text, image, and audio understanding.
Thinking Machine was founded by former OpenAI CTO Mira Murati and released its first model after a year and a half.
Inkling weights are open-sourced, and a Small preview version with 12B active parameters is provided.
Post-training data is reportedly distilled from open-source models like Kimi 2.5, but details have not been independently confirmed by the official team.
Benchmarks show Inkling is stronger than Nemotron 3 Ultra and Kimi K2.5 among open-source models, and trades wins with Kimi K2.6 and others.

Open section navigationModel Specifications and Open-Source Strategy

Model Specifications and Open-Source Strategy

Thinking Machine has released its first model, Inkling, after a year and a half in operation. It uses a MoE architecture with 975B total parameters, 41B active parameters, and a context window of 1M tokens. The model was trained on 45T tokens of data and natively supports text, image, and audio understanding.

Inkling's weights are fully open-sourced, and a Small preview version with 12B active parameters is provided. The official team states the model can be fine-tuned, and a Playground is available for user experience.

Performance Comparison: Position Among Open-Source Models

According to third-party evaluations, Inkling is stronger than Nemotron 3 Ultra and Kimi K2.5 among open-source models, and trades wins with Kimi K2.6 and DeepSeek V4 Pro: it slightly wins on IFBench, FORTRESS, and SWEBench Verified, but slightly loses on HLE, SimpleQA, and Terminal Bench.

Overall, Inkling is slightly weaker than GLM 5.2, which is stronger on multiple benchmarks such as HLE, AIME, and SWEBench Pro, and is considered the 'de facto king' of the current open-source group.

Gap and Advantages Compared to Closed-Source Models

In agentic coding, comprehensive reasoning, and factuality, Claude Fable 5 and GPT 5.6 Sol are clearly a generation ahead, for example SWEBench Verified 95% vs 77.6%, HLE-with-tools 64.5% vs 46%, SimpleQA 71.6% vs 43.9%.

However, in instruction following and not over-refusing, Inkling surpasses some closed-source models: IFBench 79.8% higher than Claude's 63.5%, FORTRESS Benign 95.9% far higher than Claude's 55.1%.

Controversy Over Distillation Data Sources

There are reports that Inkling's post-training data was distilled from a series of open-source models (e.g., Kimi 2.5). This claim currently comes only from third-party reports, and Thinking Machine has not independently confirmed specific details.

The specific proportion of distilled data and the list of source models remain unclear, which may become a focus of the open-source community, especially given that the model weights are open-sourced.

Credibility boundary

This report is based on third-party reports on X platform. Model specifications and open-source information come from official tweets, but distillation data sources and benchmark results have not been independently confirmed by the official team and should be treated as source claims rather than established facts.

Insight takeaway

The release of Inkling marks Mira Murati's team officially entering the open-source model competition. Its specifications and performance are competitive within the open-source camp, but the transparency of distillation data sources will be a subsequent focus.

Sources for this version

Primary report

歸藏(guizang.ai) (X)

Jul 16, 2026, 3:57 AM

Primary source

Same-event coverage

Also covered by 1 sources

meng shao (X)T3Thinking Machines Open-Sources Multimodal Model InklingJul 16, 2026