Breaking

Thinking Machines Lab Releases Inkling, Its First Open-Weight Model, With 975B-Parameter MoE Design and 1M-Token Context

July 15, 2026 at 16:34 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-weight foundation model, publishing the full weights on Hugging Face and offering instant access through a new Inkling Playground inside its Tinker console. The launch, announced on July 15, 2026, marks the lab's shift from providing customization tooling to shipping its own base model.

July 15, 2026 · Thinking Machines Lab

Inkling: an open-weight foundation model built to be customized, not just topped

Mira Murati's lab ships its first base model — a 975B-parameter multimodal Mixture-of-Experts with full weights on Hugging Face, betting adaptability over raw leaderboard supremacy.

975B

Total parameters
41B active per token (MoE)

45T

Pretraining tokens
text · image · audio · video

Context window
64K / 256K via Tinker API

Intelligence Index — U.S. open-weight leaders

Artificial Analysis Intelligence Index, run at maximum effort.

Inkling

Nemotron 3 Ultra

Framed as the leading result among U.S. open-weight models — ahead of NVIDIA's Nemotron 3 Ultra.

Benchmark scorecard

97.1%

AIME 2026

87.2%

GPQA Diamond

1257

Agentic Web Dev (GPT-5.6-Sol: 1260)

91.4%

VoiceBench (audio)

46.0%

Humanity's Last Exam (w/ tools; 29.7% text-only)

77.2%

MMAU (audio)

What makes it different

Controllable thinking effort — trade performance against token cost on a dial from 0.2 to 0.99.

Natively multimodal

Encoder-free: audio via dMel spectrograms, images via 40×40 patches + hMLP.

Built for Tinker

LoRA fine-tuning, agentic tool use, NVFP4 quantized weights, permissive license.

Inkling-Small

Preview variant with ~12B active parameters released alongside.

The upside

A strong new U.S. open-weight entrant — praised for efficiency and customizability, and seen as a step toward matching leading releases from China.

The caveat

Not a frontier state-of-the-art model — much of its value emerges only after fine-tuning, and the Playground saw temporary load pressure at launch.

Continue reading

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