Former OpenAI CTO Mira Murati’s AI startup, Thinking Machines Lab, has entered the race for open AI models with a release that is already drawing attention across the developer community. The company on Tuesday introduced Inkling, its first open-weights foundation model, a 975-billion-parameter Mixture-of-Experts (MoE) system built for multimodal reasoning, coding, tool use, and long-context workloads.

Early benchmark results place Inkling among the strongest open-weight models available today. It outperforms Nvidia’s Nemotron 3 Ultra on several evaluations and lands between Kimi 2.5 and Kimi 2.6 across a range of reasoning and multimodal tasks. That puts Inkling among the most capable open models released outside China and gives developers another serious alternative in a market long led by DeepSeek, Qwen, and Kimi.

The launch is a milestone for Thinking Machines Lab. Founded by Murati after leaving OpenAI, the company has spent the past year building AI systems that organizations can customize instead of relying solely on closed models. Inkling is the first foundation model the startup has trained from scratch, making it the clearest demonstration yet of its technical ambitions.

“Today, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available,” Thinking Machines said in a post on X.

Inkling: Quick Overview

Inkling combines frontier-scale architecture with practical features aimed at developers and enterprises building their own AI systems.

The model includes:

  • 975 billion total parameters, with 41 billion active per token
  • Mixture-of-Experts (MoE) architecture inspired by DeepSeek-V3
  • 1 million-token context window
  • Training on 45 trillion multimodal tokens
  • Native support for text, images, and audio
  • Controllable reasoning effort to balance quality and inference cost
  • Open weights available through Hugging Face
  • Optimized checkpoints for Nvidia Blackwell GPUs

Thinking Machines is also preparing Inkling-Small, a lighter version with 276 billion total parameters and 12 billion active parameters, offering lower inference costs for developers who do not need the full model.

Inkling is a broad, balanced generalist model. Benchmark scores are shown on a shared 0–100 scale; higher is better. The results show competitive performance across text, agentic, multimodal, and audio evaluations, rather than a model narrowly optimized for one benchmark family. This breadth reflects Inkling’s intended role: a practical multimodal foundation model for customization across domains, workflows, and products. (Source: Thinking Machines)

Built for Multimodal AI and Agentic Coding

Inkling is built as a general-purpose foundation model instead of one optimized for a single benchmark. It processes text, images, and audio natively, allowing developers to build applications without stitching together separate models for different modalities.

The model supports visual question answering, document and diagram analysis, speech transcription, spoken instruction following, and Python-based image manipulation through external tools.

Coding is another major focus. Thinking Machines says Inkling performs strongly on agentic software engineering tasks, including one-shot web application generation and longer development workflows that require planning, browser interaction, debugging, and iterative improvements. Those capabilities reflect the industry’s growing shift toward AI agents that can complete real software projects instead of generating isolated code snippets.

One of Inkling’s more distinctive features is controllable thinking effort. Developers can adjust an effort parameter between 0.2 and 0.99, allowing the model to spend more or fewer reasoning tokens depending on the task. According to Thinking Machines, Inkling matches Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly one-third as many reasoning tokens, thereby reducing inference costs without sacrificing performance.

The company also focused on calibration and factual confidence during training. Reinforcement learning based on proper scoring rules encourages the model to better estimate uncertainty rather than confidently produce incorrect answers. That work helped Inkling perform well on forecasting evaluations such as ForecastBench and Prophet Arena.

Safety received similar attention. Thinking Machines says Inkling achieves strong results on open-model safety evaluations by refusing harmful requests without rejecting large numbers of legitimate prompts.

Design Arena

Inkling was evaluated on Design Arena’s Agentic Web Dev leaderboard, where blinded human reviewers compare AI-generated web applications in head-to-head tests. The model ranks among the strongest open-weight systems on the benchmark.

Inkling’s position on Design Arena’s Agentic Web Dev leaderboard, a blinded human evaluation of generated apps. Dots refer to open-weight models. (Courtesy: Thinking Machines)

Benchmarks Put Inkling Among the Best Open Models

Inkling is positioned as a balanced generalist instead of a model built to dominate every benchmark. Across public evaluations, it consistently ranks alongside today’s strongest open-weight models.

Audio is one of its strongest categories. Inkling scores 91.4% on VoiceBench, 77.2% on MMAU, 56.6% on Audio MC, and 73.5% on MMMU Pro. On CharXiv RQ, the model reaches 78.1%, climbing to 82.0% when Python tools are enabled.

Safety benchmarks tell a similar story. Inkling posts 78.0% on FORTRESS Adversarial, 95.9% on FORTRESS Benign, and 98.6% on StrongREJECT, placing it among the strongest open-weight models currently available.

Audio and vision benchmarks against specialist omni models (open- and closed-weight), reported at effort=0.99. (Source: Thinking Machines)

Rather than chasing headline benchmark scores alone, Thinking Machines says it optimized Inkling for a combination of reasoning quality, efficiency, multimodal performance, and practical deployment.

Inkling vs. Frontier AI Models

Thinking Machines Lab compared Inkling (effort=0.99) against leading frontier AI models across a range of public benchmarks. Scores are shown as percentages unless otherwise noted. GDPVal-AA v2 is reported as a rating. Shading highlights each model’s performance relative to Inkling across all benchmarks.

Inkling vs. Frontier AI Models

Availability

Thinking Machines is releasing the full model weights through Hugging Face, including an NVFP4 checkpoint optimized for Nvidia Blackwell hardware.

Developers can fine-tune Inkling through the company’s Tinker platform or deploy it using major inference providers including Together AI, Fireworks, Modal, Databricks, and Baseten. The model is also supported by SGLang, vLLM, and Llama.cpp, TokenSpeed, and Hugging Face Transformers. An online playground lets developers experiment with the model before deployment.

Fundpluse covered Thinking Machines Lab in March after the startup announced a multi-year partnership with Nvidia that included 1 gigawatt of AI compute, significant investment, and access to next-generation Vera Rubin AI systems. That partnership gave the company access to the computing infrastructure needed to train frontier-scale models like Inkling.

Why It Matters

Inkling is more than another benchmark release. It is the first public demonstration that Thinking Machines can build frontier-scale AI models from scratch after Mira Murati’s departure from OpenAI.

The launch also adds another Western contender to an open-model landscape increasingly shaped by Chinese AI labs, including DeepSeek, Moonshot AI, and Alibaba’s Qwen. As more organizations look for AI models they can customize, fine-tune, and deploy inside their own environments, competition is shifting beyond raw benchmark performance.

Developers increasingly care about inference costs, multimodal capabilities, deployment flexibility, and the ability to adapt models to specialized workloads. Inkling arrives with strengths across each of those areas, giving enterprises, researchers, and developers another credible open alternative.

For Thinking Machines Lab, the release marks the beginning of a broader model family. For the open-source AI ecosystem, it signals that competition at the top of the leaderboard is no longer limited to a handful of established labs.