Breaking

Thinking Machines Lab Releases Inkling, a 975B-Parameter Open-Weight Multimodal Model, Topping US Open Models on a Frontend Coding Leaderboard

July 16, 2026 at 13:21 EDT

Foundation Models
Open Source
Software Dev & Coding

Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati, has released Inkling, a 975-billion-parameter open-weight multimodal model that debuted as the highest-ranked US open model on a closely watched web-development leaderboard. The company published the full weights alongside a preview of a smaller sibling model on July 15, 2026, marking its first foundation model trained from scratch.

July 15, 2026 · Thinking Machines Lab

Inkling: a 975B open-weight bet on customizable AI

Mira Murati's startup ships its first from-scratch foundation model — a multimodal Mixture-of-Experts that debuts as the highest-ranked US open model on a closely watched web-development leaderboard.

975B

total parameters

41B

active per token

45T

pretraining tokens

token context window

Sparse by design

Of 975B total parameters, only ~41B fire per token — about 1/24th of the model.

975B total

41B active

Frontend Code Arena debut

#10

among open-weight models

#37

overall ranking

1,434

arena points

The top-performing US-built open model in a top-15 field long dominated by Chinese systems like Qwen, GLM and Kimi.

Benchmark scores at max reasoning effort

Percentage-scored evaluations, drawn to scale.

91.4%

VoiceBench

78.0%

FORTRESS Adv.

77.2%

MMAU

73.5%

MMMU Pro

Strengths

Open weights in BF16 & NVFP4, fine-tunable immediately
Strong voice-task and agentic coding performance
Controllable "thinking effort" — ~1/3 the tokens of rivals on Terminal Bench 2.1

Caveats

Trails GLM 5.2 & Kimi K2.6 on HLE, Terminal Bench, SWEBench
Full BF16 weights need ~2TB of VRAM
Residual safety risks flagged in role-play and indirect prompts

Continue reading

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