Skip to main content

Ex-OpenAI CTO Murati Unveils Inkling: A New Open-Source AI Powerhouse

Former OpenAI Chief Technology Officer Miriam Murati is back in the spotlight. Her new venture, Thinking Machine Lab, just dropped its first AI model, and it's making some bold claims. Meet Inkling—a multimodal model trained from scratch that the company says is the strongest open-source AI in the United States right now.

What Makes Inkling Tick?

Inkling is no lightweight. Built on a mixture-of-experts architecture, it packs a whopping 975 billion total parameters, though only 41 billion are active at any given time. That design choice keeps things efficient while still delivering serious firepower. The model can handle up to 1 million tokens of context—enough to chew through entire books in one go—and was pre-trained on 45 trillion tokens covering four modalities: text, images, audio, and video. Best of all, the model weights are available for download, so developers can get their hands dirty.

Image

A Dream Team, But Not Unbeatable

About two-thirds of Thinking Machine Lab's core team hails from OpenAI, bringing deep expertise in cutting-edge research, product development, and AI safety. That pedigree is impressive, but it doesn't guarantee a clean sweep. When pitted against Chinese open-source models, Inkling shows some cracks, especially in reasoning and programming.

Take GLM 5.2, for instance. It leads across the board in coding, agent reasoning, and complex reasoning tasks. On SWEBench Pro, GLM 5.2 scores 62.1% versus Inkling's 54.3%. On Terminal Bench 2.1, the gap widens: 82.7% to 63.8%. DeepSeek V4 Pro also holds an edge in coding and factual tests, while Kimi K2.6 wins multiple technical benchmarks. But Inkling fights back in mathematics, scoring an impressive 97.1% on the AIME2026 test—outpacing DeepSeek.

Multimodal Prowess, Yet Still Behind the Giants

On home turf, Inkling dominates. Compared to other US open-source models like NVIDIA's Nemotron 3 Ultra, it's a clear winner in reasoning, coding, and agent workflows. But when you stack it against closed-source behemoths—Claude Fable 5, GPT-5.6 Sol, and Gemini 3.1 Pro—the story changes. These models still hold a significant lead in extreme reasoning and software engineering autonomy. Claude Fable 5 hits 95.0% on SWEBench Verified, and GPT-5.6 Sol scores 89.5 on Terminal Bench 2.1, both far beyond Inkling's numbers.

What This Means for the AI Landscape

Inkling's arrival shakes up the open-source arena, especially in the US, where it sets a new bar for multimodal capabilities. But the competition is fierce, and the gap with closed-source models reminds us that raw parameter counts and team pedigree aren't everything. For developers and researchers, Inkling offers a powerful, accessible tool—but it's not the final word in AI.

Key Points

  • Model Specs: 975B total parameters, 41B active, 1M token context, trained on 45T tokens across text, images, audio, and video.
  • Strengths: Excels in mathematics (97.1% on AIME2026) and outperforms US open-source rivals like NVIDIA's Nemotron 3 Ultra.
  • Weaknesses: Lags behind Chinese open-source models (GLM 5.2, DeepSeek V4 Pro) in reasoning and coding, and trails closed-source giants (Claude Fable 5, GPT-5.6 Sol) in extreme reasoning and software engineering.
  • Availability: Model weights are open-source and downloadable.