Inkling
An open-weights mixture-of-experts model handling text, images and audio input
Inkling is a large language model from Thinking Machines, released on 15 July 2026 under an Apache 2.0 licence. It accepts text, pictures and sound as input and is built for chat, coding, multilingual work and agentic systems, with a reasoning budget that users can tune per task.
- First large language model released by Thinking Machines, on 15 July 2026
- 975 billion parameters, with 41 billion active per token
- Context window of 1,048,576 tokens
- Trained on 45 trillion tokens of text, image, audio and video
- Reasoning budget adjustable from 0.2 to 0.99
What it is
Inkling is a large language model put out by Thinking Machines on 15 July 2026. Weights are published under Apache 2.0, so anyone can download and run it. The system reads text, images and audio, and replies in text, including code and structured data. It is aimed at chat, software development, multilingual use and agentic systems. Inside, it holds 975 billion parameters, though only 41 billion of those are used for any single token. That mixture-of-experts arrangement keeps the computing cost down while keeping capacity high.
How it came about
This was the first large language model that Thinking Machines released. It arrived on 15 July 2026 with an Apache 2.0 licence attached to the weights. Training drew on 45 trillion tokens spanning text, images, audio and video. The company also runs Tinker, a platform where developers can fine-tune the model for their own purposes.
Inside the model
Inkling uses a mixture-of-experts design: incoming material is routed to whichever internal parts are best suited to it, so computing power is spent where it helps. The context window reaches 1,048,576 tokens. Training covered 45 trillion tokens across four media types, and the model reasons natively over any of them. Output is text only. A knob lets users set the reasoning budget, the amount of internal deliberation before a reply, anywhere from 0.2 to 0.99, matching effort to the job. Deployment works with tools such as llama.cpp, and the hardware floor is 2 TB of VRAM, or 600 GB for a quantised build.
What it can do
The model handles chat, coding, multilingual tasks and agentic systems, taking text, images and audio as input and producing text, including source code and structured data. Its reasoning effort can be dialled up or down per request through the budget setting, which lets callers trade cost against depth. It reasons natively across all four input formats it was trained on.
How to use it
Weights are downloadable under an Apache 2.0 licence. The model runs through common inference libraries such as llama.cpp, and can also be served by third-party providers. Fine-tuning for a specific job is available through Tinker, the Thinking Machines platform. A quantised build lowers the VRAM requirement from 2 TB to 600 GB.
Safety and testing
Inkling was deliberately built to resist censorship and give direct answers on topics that are politically sensitive or blocked elsewhere. Its makers advise handling content moderation with outside tools rather than relying on the model's own safety checks, which they say can still let harmful requests through.
Questions people ask about Inkling
What is Inkling?
It is an open-weights large language model from Thinking Machines, released on 15 July 2026 under Apache 2.0. It takes text, images and audio as input and answers in text, including code and structured data, for chat, development, multilingual work and agentic systems.
How large is it?
The model holds 975 billion parameters in total. Because it uses a mixture-of-experts design, only 41 billion of those are active for any given token, which keeps computing use efficient while retaining capacity.
What hardware do I need to run it?
Running it normally asks for at least 2 TB of VRAM. A quantised version exists that brings that minimum down to 600 GB. It can be deployed with common inference libraries such as llama.cpp.
Can I fine-tune it?
Yes. Developers can adapt it to a particular job through Tinker, the platform from Thinking Machines, and can serve the result via third-party providers.
How does it handle sensitive topics?
It was built to resist censorship and give direct answers on politically sensitive subjects. Its creators recommend external moderation tools, saying internal safety checks in language models do not always stop harmful prompts being answered.
Good reading and sources
The facts on this page come from the Wikipedia article "Inkling (large language model)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.