Breaking

Thinking Machines Lab Debuts First Open-Weight Model 'Inkling' Under Apache 2.0 License

July 16, 2026 at 23:42 EDT

Foundation Models
Open Source

Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati, has released its first model, an open-weight multimodal system called Inkling, freely available under the Apache 2.0 license. The launch marks the lab's public entry into the increasingly crowded market for large language models, but with a deliberately different bet than the closed offerings that dominate consumer AI.

Thinking Machines Lab · First Model

Inkling: an open-weight bet against one-size-fits-all AI

Mira Murati's startup debuts a multimodal Mixture-of-Experts model under Apache 2.0 — prioritizing customization, cost control and data sovereignty over chasing frontier leaderboards.

975B

Total parameters
~41B active per token

Token context window

45T

Training tokens
text · image · audio · video

256

Experts (6 + 2 shared fire per pass)

Hardware to run the full model

VRAM required at each precision — quantization to NVFP4 cuts the footprint by ~70%.

~2TB

BF16 precision
≈ 8× B300 / 16× H200

~600GB

NVFP4 quantized

Benchmarks at maximum thinking effort

Competitive among open-weight systems, though generally trailing leading closed models.

SWE-Bench Verified77.6%

VoiceBench91.4%

MMMU Pro73.5%

Terminal Bench 2.163.8%

Humanity's Last Exam (text)29.7%

Developers welcome it

Meaningful step for open AI
Strong base for fine-tuning via Tinker
Efficiency-first, customizable, cost-controlled

But with caveats

Trails rivals like GLM 5.2 on some benchmarks
Sheer size makes local execution hard
Hallucination, bias, long-context degradation

Continue reading

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