Image: Thinking Machines Lab / thinkingmachines.ai Mira Murati's First Model Is Out. It's Not Trying to Beat Fable 5.
Thinking Machines Lab released Inkling, a 975B-parameter open-weight model trained on 45 trillion tokens across text, image, audio, and video. Former OpenAI CTO Mira Murati is betting enterprises want AI they can customize, not just rent.
Mira Murati left OpenAI in September 2024, took nine months, and on July 15 shipped a 975-billion-parameter model trained on 45 trillion tokens. Thinking Machines Lab calls it Inkling. The company is explicit that it’s not the strongest model available — open or closed. That’s part of the pitch.
The Architecture
Inkling is a mixture-of-experts system. Total parameters: 975 billion. Active parameters per forward pass: 41 billion — about 256 routed experts per layer with 6 active per token. It uses a combination of sliding-window and global attention layers in a 5:1 ratio, with a 1 million token context window.
The training data spans text, images, audio, and video — 45 trillion tokens total. Current outputs are limited to text, code, styled artifacts, and structured data. The model uses relative positional embeddings rather than RoPE, and incorporates short convolutions for long-context efficiency.
Post-training involved over 30 million RL rollouts. Reasoning performance improved log-linearly throughout.
The Benchmarks
On SWE-bench Verified, Inkling scores 77.6% — which puts it in genuine contention with frontier coding models. AIME 2026: 97.1%. GPQA Diamond: 87.2%. HLE with tools: 46%.
The token efficiency finding is the interesting one. On some coding benchmarks, Inkling achieves equivalent results to Nvidia’s Nemotron 3 Ultra at roughly one-third the tokens. A Bridgewater Associates pilot with a fine-tuned version scored 84.7% on financial reasoning — beating proprietary models while costing about one-fourteenth as much to operate.
Thinking Machines also previewed Inkling-Small, a 276B-parameter variant with 12B active parameters that matches or exceeds the larger model on many benchmarks.
Why Open-Weight
The full weights are on Hugging Face, including an NVFP4 checkpoint for NVIDIA Blackwell hardware. Inkling is available for fine-tuning via Tinker (Thinking Machines’ customization platform, currently at a 50% discount), and via APIs on TogetherAI, Fireworks, Modal, Databricks, and Baseten. There’s also a free playground.
The reasoning behind open-weight release isn’t altruism. Thinking Machines’ revenue comes from Tinker, not from metered API access. Releasing the weights expands the base of developers who might want to customize — and customizing requires Tinker.
The company’s stated thesis: AI that organizations can adapt for themselves will outperform one-size-fits-all models from the major labs. Inkling is meant to be infrastructure, not a finished product.
Nine Months
The part that stands out is the timeline. OpenAI took roughly five years to reach this capability level. Anthropic took about three. Thinking Machines did it in nine months with around 200 people after some early co-founder departures, including two who left for OpenAI in January.
Inkling is not claiming to be the benchmark leader. On safety evaluations — FORTRESS Adversarial (78%) and StrongREJECT (98.6%) — it performs well. It was also trained for calibrated uncertainty, which means it’s designed to acknowledge what it doesn’t know rather than confabulate.
Whether the enterprise customization bet pays off is a separate question. But the model exists, the weights are available, and Murati got it done faster than almost anyone expected.
Sources: Thinking Machines Lab, TechCrunch, Axios