Breaking

Thinking Machines Lab Releases Inkling, Mira Murati's First Open-Weights Foundation Model

July 15, 2026 at 15:18 EDT

Foundation Models
Open Source
AI Agents

Thinking Machines Lab, the AI startup led by former OpenAI CTO Mira Murati, has released Inkling, its first open-weights foundation model, positioning it as a customizable alternative to closed frontier systems rather than an outright performance leader. The model launched on July 15, 2026, with full weights published on Hugging Face and immediate fine-tuning available through the company's Tinker platform.

July 15, 2026 · Thinking Machines Lab

Inkling: An Open-Weights Bet on Customization Over Raw Power

Mira Murati's startup releases its first foundation model trained from scratch — a natively multimodal Mixture-of-Experts system with full weights on Hugging Face and instant fine-tuning via Tinker. It is framed as a starting point to build on, not a benchmark champion.

975B

Total parameters (MoE)

Token context window

45T

Pre-training tokens
(text · image · audio · video)

Sparse by design: only a fraction runs at inference

Of 975B total parameters, roughly 41B are active per pass — the Mixture-of-Experts routing that keeps a huge model efficient to serve.

975B

total parameters

41B

active at inference

Controllable reasoning effort

Users dial inference-time compute up or down. On tests like Terminal Bench, Inkling reports comparable performance while spending far fewer tokens — in some cases about a third of what a rival consumed.

~3×

rival token use

~1×

Inkling tokens · similar quality

What ships with the release

Full weights

Open on Hugging Face · fine-tune instantly via Tinker

NVFP4 build

Optimized for efficient inference on NVIDIA Blackwell

Inkling-Small

Preview variant · ~12B active parameters

Developer optimism

A strong open base that's easy to customize and pairs with Tinker for rapid fine-tuning — eyed for agentic coding, web app generation and multimodal reasoning.

The caveat

The company itself says it isn't the most capable model on the market. Days-old and largely untested externally, its true standing awaits independent verification.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.