# Thinking Machines Lab debuts Inkling, its first open model

> The lab's inaugural open-weights release is a mixture-of-experts system that takes image and audio inputs, shipped under a permissive Apache 2.0 license.

Published by The Open Weights on Jul 15, 2026. Canonical: https://theopenweights.com/news/inkling-wh8c

## Key facts

- Company: Thinkingmachines
- Model: Inkling
- Version: Inkling
- Category: Any-to-Any
- Modalities: Any-to-Any, Vision-Language, Text / LLM
- License: Apache 2.0 (Open weights, commercial use allowed)
- Context window: 1.0M tokens
- Architecture: InklingForConditionalGeneration
- Precision: BF16
- Weights size: 1904.8 GB
- Significance: major
- Published: 2026-07-15
- Last verified: 2026-08-17
- Hugging Face: https://huggingface.co/thinkingmachines/Inkling
- Canonical URL: https://theopenweights.com/news/inkling-wh8c

Thinking Machines Lab has released Inkling, described as its first open-weights model and a mixture-of-experts (MoE) multimodal system that accepts image and audio inputs alongside text. The model is published under the permissive Apache 2.0 license, according to the lab's [announcement](https://thinkingmachines.ai/news/introducing-inkling/).

The MoE design is notable because it activates only a subset of the model's parameters for any given input, a routing approach that lets teams scale total capacity while keeping the compute cost of each forward pass in check. Pairing that architecture with native image and audio handling puts Inkling in the growing category of open models built to reason across more than one input type.

## Why it matters

An Apache 2.0 license is one of the least restrictive terms an AI lab can choose, which means developers can study, fine-tune, and deploy Inkling commercially without the usage caveats attached to many so-called open releases.

- **Open weights**: the model parameters are available for download and self-hosting.
- **Multimodal inputs**: image and audio are supported in addition to text.
- **MoE architecture**: sparse expert routing aimed at efficiency at scale.

As an initial release, Inkling establishes a baseline rather than iterating on a prior version. The full picture — including parameter counts, context length, and benchmark results — will come into focus as the community begins testing the weights against comparable open multimodal systems.

## Get the model

- [Hugging Face](https://huggingface.co/thinkingmachines/Inkling)
- [Hacker News](https://thinkingmachines.ai/news/introducing-inkling/)
- [Announcement](https://huggingface.co/blog/thinkingmachines-inkling)

## Sources

- [Inkling: Our Open-Weights Model](https://thinkingmachines.ai/news/introducing-inkling/) — Hacker News, Jul 15, 2026
- [thinkingmachines/Inkling](https://huggingface.co/thinkingmachines/Inkling) — Hugging Face, Jul 14, 2026
- [Welcome Inkling by Thinking Machines](https://huggingface.co/blog/thinkingmachines-inkling) — Announcement, Jul 15, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "Thinking Machines Lab debuts Inkling, its first open model", The Open Weights, Jul 15, 2026, https://theopenweights.com/news/inkling-wh8c