On November 12, 2026, Thinking Machines Lab published the weights of Inkling 2 Small on Hugging Face. It is the company's second-generation model, first presented on November 5. The "weights" are the model itself: having them means you can download it, run it on your own servers and adapt it, without having to go through a paid service.
The license is Apache 2.0, one of the most permissive: it also allows commercial use and integration into your own products. That puts Inkling 2 Small on the growing list of open models that aim to compete with the closed ones from the big providers.
What is Inkling 2 Small, in practice?
It is an MoE model (Mixture of Experts): inside it there are many specialized parts, but only a small fraction of them work on each word. That makes the model huge on paper but lighter to run. It reads text and images, answers with text and lets you adjust how much it "thinks" before answering. These are the main figures stated by the company:
| Feature | Inkling 2 Small |
|---|---|
| Parameters | 594 billion in total, 17 billion active per word |
| Input | Text and images (no audio) |
| Context | Up to 1 million tokens |
| License | Apache 2.0 |
| Weight formats | NVFP4 (recommended) and MXFP8 |
| Where to use it | Hugging Face, the Tinker platform and APIs from NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare |
How good is it, according to the stated tests?
The company presents it as designed for agents, coding, research and office work. Some results stated by Thinking Machines Lab (not independently verified):
- Document reading (OmniDocBench v1.6): 94.6%;
- Understanding images and charts (MMMU Pro): 81.6%;
- Coding (DeepSWE v1.1): 72.1%;
- Web research (BrowseComp, with context compaction): 90.2%;
- Artificial Analysis Intelligence Index: about 31, a clear jump over previous Inkling models according to the company.
The announcement also lists the limits: less concise answers than earlier versions, sometimes picking the wrong tool when many are available, less polished presentation in conversations and no audio support. On alignment tests, meaning how well it follows behavior rules, the company itself says it still trails OpenAI's GPT-6.1 Sol.
Can you run it on your own computer?
Not on a regular PC. According to the official model card, the NVFP4 version needs at least 500 GB of combined video memory (for example 2 NVIDIA B300 or 4 H200 cards); the MXFP8 version at least 800 GB. The Hugging Face page also shows compressed versions for llama.cpp, Ollama and LM Studio, and the model is listed in Ollama's library, but with 594 billion parameters these are still files of hundreds of gigabytes, suited to servers or very powerful workstations.
If you want to know which models really run on an office computer, read our guide running AI locally: what it is, when it makes sense and what you need (in Italian), and our article on llama.cpp 1.0 (in Italian), the engine behind many of these programs.
What changes for people who work with AI?
For an individual professional, not much right away: you will mostly try the model through the API providers listed above. For a company with a technical team, though, it is one more option: an open model with a commercial license and a very long context window, which can be installed on your own servers to process documents, contracts or code without sending data to third parties.
As with any model, the official card recommends human oversight: it can make things up ("hallucinations") and should not be used in medical, legal or safety-critical settings without further adaptation and validation. The test numbers are the ones stated by the company: before choosing it, test it on your own real cases.
Frequently asked questions
Is Inkling 2 Small free?
The weights can be downloaded for free from Hugging Face under the Apache 2.0 license. If you use it through an API (Tinker or the other providers), you pay per use according to each service's pricing.
Why is it called "Small" if it has 594 billion parameters?
The name places it within the Inkling 2 family. For users, the number of active parameters, 17 billion, matters more, because it affects the speed and cost of each answer.
Does it handle languages other than English?
The official card publishes no per-language figures and warns that performance may vary across languages and domains. Before using it for texts in other languages, run a few tests.
Official sources
- Thinking Machines Lab: Inkling 2 Small announcement
- Thinking Machines Lab: model card
- Hugging Face: thinkingmachines/Inkling-2-Small
- Ollama: inkling-2-small
This article is written with the help of artificial intelligence models, based on the official sources listed, and is reviewed before publication.














