Inkling
Thinking Machines' first open model: a 975B multimodal mixture-of-experts under Apache 2.0, built as a base for customization.
In short. Inkling is an AI model that understands text, pictures and spoken audio, released with full weights under the Apache 2.0 license by Thinking Machines Lab. The company says it is not trying to top every leaderboard; it is meant to be a good starting point that you retrain for your own needs. A smaller Inkling-Small is also available.
Summary
Inkling was released on 15 July 2026 by Thinking Machines Lab, the AI lab founded by former OpenAI CTO Mira Murati. It is the lab's first production language model and the first in a planned family. A smaller sibling, Inkling-Small, followed on 27 July 2026.
Inkling is a decoder-only mixture-of-experts transformer with 975 billion parameters in total and 41 billion active per token. It has 66 layers; each token is routed to 6 of 256 experts, plus 2 shared experts that run on every token, and attention alternates between local and global layers. It was pretrained on 45 trillion tokens of text, images, audio and video. Inputs can be text, images and audio (WAV, up to about 20 minutes); output is text. The open weights support a context window of up to 1 million tokens, while the lab's own hosted service offers 64K and 256K options.
Inkling-Small has 276 billion parameters and 12 billion active, with the same architecture. Both come as BF16 checkpoints and as NVFP4 versions for NVIDIA Blackwell GPUs, and include multi-token-prediction layers for faster generation. Thinking Machines states plainly that Inkling is not the strongest overall model, open or closed. Its pitch is multimodality, efficient thinking and easy fine-tuning, including on the lab's Tinker platform. The license is Apache 2.0.
Good for
- A base model to fine-tune for your own data, including audio tasks
- Voice and image assistants that need one model for text, images and speech input
- Companies that want Apache 2.0 terms for a US-built frontier-scale model
- Inkling-Small for teams that want the same design on fewer GPUs
Watch out for
- The full model needs about 1.9 TB in BF16 or about 0.6 TB in NVFP4
- The developer itself says it is not the top performer overall; it is a customization base
- It outputs text only, so spoken replies need a separate text-to-speech model
- Check each checkpoint, because Inkling-Small is the more practical size for most teams
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Inkling-Small NVFP4 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Inkling-Small BF16 | 276B 12B active | 635 GB | 317 GB | 190 GB | Big-memory desktop |
| Inkling-Small NVFP4 NVFP4 | 276B 12B active | 635 GB | 317 GB | 188 GB | Big-memory desktop |
| Inkling BF16 | 975B 41B active | 2.24 TB | 1.12 TB | 673 GB | GPU server |
| Inkling NVFP4 NVFP4 | 975B 41B active | 2.24 TB | 1.12 TB | 653 GB | GPU server |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source LLMs in 2026
- Open models that listen: nine take audio, but only three answer in speech
- Giant open models now ship in 4-bit: what it means for memory planning
Frequently asked questions
- What is Inkling?
- Inkling is an AI model that understands text, pictures and spoken audio, released with full weights under the Apache 2.0 license by Thinking Machines Lab. The company says it is not trying to top every leaderboard; it is meant to be a good starting point that you retrain for your own needs. A smaller Inkling-Small is also available.
- Can I use Inkling commercially?
- Yes. Inkling is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Inkling open source?
- Inkling is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Inkling on my own computer?
- Only on a machine with a lot of memory. The smallest variant, Inkling-Small NVFP4, needs about 188 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was Inkling released?
- Inkling was released by Thinking Machines Lab on July 15, 2026.
- How long a context window does Inkling have?
- Inkling supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Inkling: Our Open-Weights Model (Thinking Machines Lab, 15 Jul 2026) thinkingmachines.ai
- Inkling model card (Thinking Machines Lab) thinkingmachines.ai
- Welcome Inkling by Thinking Machines (Hugging Face blog) huggingface.co
- Inkling on Artificial Analysis (15 Jul 2026) artificialanalysis.ai
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Muse Spark Open Meta | 816B 42B active | 1M | Muse Model License (Conditional) | 563 GB GPU server | |
| Mistral Large 4 Mistral AI | 1.05T 52B active | 1M | Mistral Open-Weight (Conditional) | 671 GB GPU server | |
| Reflection Beam Reflection AI | 501B 23.4B active | 1M | Apache 2.0 (Permissive) | 323 GB Big-memory desktop | |
| Arete-27B Nous Research | 27.4B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU |