Breaking

Thinking Machines Lab Releases Inkling, Its First Open-Weights Model, Drawing on Chinese Rivals' Efficiency Playbook

July 15, 2026 at 15:12 EDT

Foundation Models
Open Source

Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati, has published Inkling, its debut model and its first open-weights release, borrowing design ideas from a wave of efficient Chinese AI systems. The company unveiled the model on July 15, 2026, releasing the full weights alongside a smaller preview variant and a self-service fine-tuning path for developers.

July 15, 2026 · Thinking Machines Lab

Inkling: Mira Murati's Startup Ships Its First Open-Weights Model

A natively multimodal Mixture-of-Experts transformer built for efficiency — borrowing design ideas from a wave of fast-moving Chinese open-weights labs, and betting on customization over chart-topping benchmarks.

~975B

total parameters (MoE)

~41B

active per token

token context window

45T

tokens pre-trained on

The efficiency claim, drawn to scale

Matches Nemotron 3 Ultra on Terminal Bench 2.1 — using ~⅓ the generated tokens

Same score, far fewer tokens spent reasoning. Each block ≈ one-third of the tokens.

~3× tokens

Nemotron 3 Ultra

~1× tokens

Inkling

Reported benchmark scores

VoiceBench91.4%

SWE-Bench Verified77.6%

MMMU Pro73.5%

What developers like

Genuinely competitive open-weights debut
Easy fine-tuning via the Tinker platform
Multimodal foundation: text, image, audio
Permissive Apache 2.0-equivalent license

The caveats

Not the outright leader on every benchmark
BF16 weights need >2TB VRAM
NVFP4 build still needs ~600GB
Numbers await independent reproduction

A two-way current across the Pacific

Inkling borrows efficiency and multimodal ideas from Chinese open-weights systems like GLM-5.2, Qwen3-Omni and the Kimi family — while labs such as Yoolee and InfiX.ai reportedly move to challenge Thinking Machines Lab with their own industry-focused efforts.

Continue reading

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