Breaking

Thinking Machines Lab Releases Inkling, Its First Open-Weight Foundation Model Under Apache 2.0

July 16, 2026 at 08:02 EDT

Foundation Models
Open Source
AI Agents

Thinking Machines Lab, the AI company founded by former OpenAI chief technology officer Mira Murati, released its first foundation model on July 15, 2026, publishing the full weights of a natively multimodal system called Inkling under a permissive Apache 2.0 license. Unlike the closed, proprietary systems that power ChatGPT and Claude, Inkling ships as an open-weights model that developers can download, inspect, run locally, and fine-tune without restriction.

July 15, 2026 · Thinking Machines Lab

Inkling: An Open Bet Against One-Size-Fits-All AI

Mira Murati's Thinking Machines Lab releases its first foundation model — a natively multimodal Mixture-of-Experts system shipped with full open weights under Apache 2.0, built for teams that value control, transparency, and cost over top-line benchmark scores.

975B

total parameters (MoE)

41B

active parameters per token

token context window

45T

training tokens across 4 modalities

Only a Sliver Runs at a Time

Of 975B total parameters, roughly 41B are active per token — a ~24× sparse MoE design that keeps compute low relative to model size.

41B

active

975B

total

Architecture at a Glance

EXPERTS

256 experts · 6 routed + 2 shared per token

INPUTS

Text · image · audio (16kHz, ~20 min) — encoder-free

FORMATS

BF16 · MXFP8 · NVFP4 + multi-token prediction

THINKING EFFORT

Tunable 0.2–0.99: trade compute for performance

Quantization Halves the Hardware

NVFP4 quantization cuts the GPU footprint roughly in half versus full-precision BF16.

BF16 (full precision)

~8× B300 or 16× H200

NVFP4 (quantized)

~4× B300 or 8× H200

THE CASE FOR

Strong open-weight base for coding & agentic work
Native multimodal reasoning across voice and vision
Token-efficient — matches rivals with ~⅓ the tokens in some cases
Easy to fine-tune; adjustable compute for cost control

THE CAVEATS

At max performance, still trails leading closed systems
Safety evaluations land around the open-weight average
A versatile generalist — not an outright benchmark leader

Available From Day Zero

Hugging Face Tinker

Continue reading

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