AI models

Inkling 2 Small goes open weight: Thinking Machines' model is ready to download

On November 12, 2026 Thinking Machines Lab, the company founded by Mira Murati, published the weights of Inkling 2 Small on Hugging Face under an Apache 2.0 license. It is a 594-billion-parameter model that reads text and images, launched on November 5 on the Tinker platform only. Here is what it contains, what it costs to use, what it takes to run and what it changes for you.

Published on 7 min readBy the editorial team of The AI Creator System

Leggi questo articolo in italiano

AI-generated cover: a cyan and violet crystal pen nib pouring glowing ink onto an open padlock, with the title Inkling 2 Small now open weight
Cover image generated with artificial intelligence. How we use AI

Inkling 2 Small open weights: what was released on November 12

Inkling 2 Small is Thinking Machines Lab's second-generation model, and since November 12, 2026 its weights are free to download. Thinking Machines Lab is the research company founded by Mira Murati, OpenAI's former chief technology officer. The model was announced on November 5, 2026 on Thinking Machines' official blog, at first only through Tinker, the company's platform for using and training models, with closed weights.

In that same announcement the company promised to release the full weights "next week", after external safety testing and red-teaming. It kept the promise: the post was updated on November 12, 2026 to say the weights are available on Hugging Face. The thinkingmachines/Inkling-2-Small repository lists the apache-2.0 license, and a second repository holds the MXFP8 version.

Austrian tech outlet Trending Topics covered the launch on November 6, 2026, noting that the weights date was not yet confirmed at the time. Now it is.

Inkling 2 Small specs at a glance

Inkling 2 Small is a Mixture-of-Experts model: it has 594 billion parameters but uses about 17 billion of them for each token. Here are the figures from the official model card:

ItemInkling 2 Small
Total parameters594 billion
Active parameters per token17 billion (Cloudflare lists 18)
Experts384, with 8 routed per token
Contextup to 1 million tokens
Input / outputtext and images / text only
Audionot supported
LanguagesEnglish, with general multilingual support
Weight formatsNVFP4 (recommended) and MXFP8
LicenseApache 2.0

The model was trained directly in the compressed NVFP4 format, which the card calls its native precision. According to Thinking Machines, the heavier MXFP8 version gains only 1-2 percentage points on evaluations. Images go through a 292-million-parameter vision encoder. Unlike the first Inkling generation, this version does not handle audio.

License, pricing and hardware: how to use Inkling 2 Small

The Apache 2.0 license lets you use, modify, fine-tune and sell products built on Inkling 2 Small, including for commercial purposes. It is more permissive than the license of Mistral Large 4, which requires a separate agreement for anyone reselling the model as an API. Thinking Machines' Model Acceptable Use Policy still applies.

There are three ways to use it:

  • Third-party APIs. According to the official blog it is available through NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. On Cloudflare Workers AI it costs 0.30 dollars per million input tokens, 1 dollar per million output tokens and 0.10 dollars for cached tokens, on the Workers Paid plan.
  • Tinker. Thinking Machines' platform offers a playground to try it and tools for fine-tuning.
  • Downloaded weights. You can run them with SGLang, vLLM, TokenSpeed, Unsloth or the Hugging Face library.

Running it yourself, though, takes data-center hardware. The NVFP4 version needs at least 500GB of total GPU memory, for example 2 NVIDIA B300 or 4 H200 GPUs; the MXFP8 version needs at least 800GB, meaning 4 B300 or 8 H200. On Cloudflare the model also offers six reasoning levels, from none to max, to balance speed and quality.

Inkling 2 Small benchmarks: what Thinking Machines claims

The published results come from Thinking Machines and should be read as claims: different tests use different setups. The main figures claimed in the official announcement:

  • 31 points on the Artificial Analysis Intelligence Index (30.6 in the table), up from 13 for its predecessor Inkling Small;
  • 96.5% on ARC-AGI-1 and 81.7% on ARC-AGI-2, reasoning tests on unfamiliar problems;
  • 81.6% on MMMU Pro and 83.8% on ScreenSpot Pro, tests of image and screen understanding;
  • a top placing on DesignArena, a leaderboard of AI-generated web pages and interfaces.

According to Trending Topics, only two Chinese open models, DeepSeek V4.1 Flash and GLM-5.3, rank above Inkling 2 Small on the Artificial Analysis index with 32 points, while Mistral Large 4 scores 26. The leading closed models remain far ahead: the same outlet cites 58 points for Claude Opus 5.5. In short, Inkling 2 Small is among the strongest open models built outside China, not a rival to the most powerful closed models.

Inkling 2 Small limitations, according to the company

Thinking Machines lists Inkling 2 Small's weak points itself, and they are worth knowing before you pick it. According to the official announcement:

  • it is not the most concise model and tends toward long answers;
  • it does not always choose the best tool, especially in codebases with many skills;
  • it does not support audio;
  • it is less polished as a chat assistant in how it organizes and presents content;
  • it overreacts to mental health and safety-related questions;
  • on alignment evaluations it still trails closed models such as GPT-6.1 Sol.

The model card adds the usual limitations: hallucinations, occasionally imprecise instruction-following and weaker performance in long conversations. It advises against medical, legal or safety-critical use without further fine-tuning, validation and human oversight.

What it means for you: creators, freelancers and small businesses

For creators and small business owners, an open Inkling 2 Small means one more low-cost option through an API, not a model to install in your office. Here is how to use it sensibly:

  1. Try it through an API. At 0.30 and 1 dollar per million tokens you can run plenty of tests for a few dollars: summaries of long documents, analysis of screenshots and charts, text drafts.
  2. Use the vision side. Thinking Machines highlights image interpretation and interface creation: you can ask it to read a screenshot, a table or a web page draft.
  3. Check non-English output. The card lists English as the main language with general multilingual support, so compare its results with Claude, ChatGPT and Gemini on your own texts before relying on it.
  4. Do not self-host it. Without a server with at least 500GB of GPU memory, the API makes more sense. For open models that really run on a computer, look at Nemotron 3.5 or Liquid AI LFM3.

The method matters more than the model: well-written prompts get good results from any large model. The 80 prompts in Claude Business OS cover email, meetings, procedures, documents and numbers; they are written for Claude, but they also work well for testing Inkling 2 Small on office tasks. If you want to use its vision skills to build pages and websites, start with Website AI University. To find your way among the tools for creating and selling digital products, there is the AI Business Toolkit.

For creative AI tools, start with our AI guides. For a complete path there is The AI Creator System, and here is how we use artificial intelligence.

Frequently asked questions

When were the Inkling 2 Small weights released?

On November 12, 2026, on Hugging Face, in the thinkingmachines/Inkling-2-Small repository. The model had launched on November 5, 2026, at first only on the Tinker platform.

Is Inkling 2 Small free?

The weights are a free download and the Apache 2.0 license also allows commercial use. Using it through an API costs money: on Cloudflare Workers AI, 0.30 dollars per million input tokens and 1 dollar per million output tokens.

Can I run Inkling 2 Small on my own computer?

Realistically, no: the recommended version needs at least 500GB of GPU memory, for example 2 NVIDIA B300 or 4 H200 GPUs. For everyday use, the API makes more sense.

Does Inkling 2 Small understand images?

Yes: it accepts text and images as input, but it only answers with text. It does not generate images and does not support audio.

Who made Inkling 2 Small?

Thinking Machines Lab, the artificial intelligence research company founded by Mira Murati, OpenAI's former chief technology officer.

Read next

Sources checked on November 16, 2026