Inkling AI is the family of open-weights models from Thinking Machines Lab, the startup Mira Murati founded in February 2025 after leaving OpenAI. There are three models so far: Inkling (July 15, 2026), Inkling-Small (July 30) and Inkling 2 Small, announced on November 5 and released as open weights on November 12. The newest one scores 31 on the Artificial Analysis Intelligence Index v5.0.1. The original Inkling, a much bigger model, scores 12 on the same index.
Most pages about Inkling were written in launch week in July, before the second generation existed and before Artificial Analysis rebuilt its index. That is why a score of 41 still circulates for the first model, and why some pages still quote the wrong parameter counts. This guide covers all three models with current numbers: what each one is, how good it is, what it costs, how to try it for free, and whether your Mac can run it.
The Key Takeaways
- Three models, one family: Inkling (975B parameters, 41B active), Inkling-Small (276B, 12B active) and Inkling 2 Small (594B, 17B active), all under the Apache 2.0 license.
- The new one is the good one: Inkling 2 Small scores 31 on Artificial Analysis’ Intelligence Index v5.0.1, against 12 for Inkling and 13 for Inkling-Small.
- Cheap to use, free to try: Inkling 2 Small costs $0.30 in and $1.00 out per million tokens on Tinker and OpenRouter, and OpenRouter lists a free variant.
- No consumer app: the only first-party chat window is the developer-facing Tinker Playground.
- Running it at home is hard: the smallest Inkling 2 Small download is 171.7GB, more memory than any MacBook has.
What Is Inkling AI?
Inkling AI is a family of open-weights, multimodal AI models made by Thinking Machines Lab and released under the Apache 2.0 license. Open weights means anyone can download the trained model and run it on their own hardware. They can also fine-tune it and build commercial products on it, with no permission or license fee needed.
Thinking Machines Lab was founded in February 2025 by Mira Murati, OpenAI’s former chief technology officer, with a group of other former OpenAI staff. According to Fortune, the startup raised $2 billion at a $12 billion valuation in 2025, before it had shipped anything. Its first product was not a model at all. Tinker, launched on October 1, 2025, is a platform for fine-tuning other companies’ open models. Inkling, released on July 15, 2026, is the first model Thinking Machines trained from scratch.
Thinking Machines is modest about it. The launch post says Inkling “is not the strongest overall model available today, open or closed.” The pitch is that Inkling is a good base to customize: broad across text, images and audio, efficient with its thinking, and ready to fine-tune on Tinker the day it ships. Thinking Machines also built it to serve as the background reasoning model for its “interaction models”, a real-time voice and vision system the lab had introduced shortly before.
Every Inkling model is a mixture of experts. Only a slice of the model, the “active” parameters, does the work on each token. Each token of a 594B-parameter model therefore takes about as much compute as a 17B model would. That is why Tinker prices mixture-of-experts models by active parameters rather than total size, and why a model with 594B parameters can cost less to run than one with 276B.
Thinking Machines Inkling Models Compared
Thinking Machines has released three Inkling models in four months, and the newest one is the best by a wide margin. The table compares them on the facts that matter when choosing one.
| Inkling | Inkling-Small | Inkling 2 Small | |
|---|---|---|---|
| Released | July 15, 2026 | July 30, 2026 | Nov 5, 2026 (weights Nov 12) |
| Total parameters | 975B | 276B | 594B |
| Active parameters | 41B | 12B | 17B |
| Context window | Up to 1M tokens | Up to 1M tokens | Up to 1M tokens |
| Inputs | Text, image, audio | Text, image, audio | Text, image |
| AA Intelligence Index v5.0.1 | 12 | 13 | 31 |
| Tinker serverless price (per 1M in / out) | $1.00 / $4.05 | $0.30 / $1.20 | $0.30 / $1.00 |
| License | Apache 2.0 | Apache 2.0 | Apache 2.0 |
Two numbers on other sites need care. Thinking Machines gives Inkling as 975B parameters; the 952B some pages quote is the tensor count of the files on Hugging Face, not the lab’s figure. Inkling-Small has 276B parameters, of which 12B are active; it is not a 12B model. The 1M-token context is what the weights support. Tinker serves all three models at 64K or 256K.
Inkling, the July original
Inkling is the largest model in the family, with 975B parameters and 41B active. It has 66 layers, and each token is routed to 6 of 256 experts plus 2 shared ones. Thinking Machines pretrained it on 45 trillion tokens of text, images, audio and video. Inkling reads text, images and audio and writes text. Thinking Machines’ headline efficiency claim is that Inkling matches Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the tokens.
Inkling-Small, a quarter of the size
Inkling-Small arrived two weeks later with 276B parameters and 12B active, and Thinking Machines says it achieves comparable performance to Inkling at a quarter of the size. It started training after Inkling, so it got an improved recipe. Thinking Machines reports that Inkling-Small beats Inkling on reasoning and agentic coding: 31.6% on Humanity’s Last Exam against Inkling’s 29.7%. Inkling keeps the edge on knowledge and factuality.
Inkling 2 Small, the second generation
Inkling 2 Small is the first model of Thinking Machines’ second generation, with 594B parameters and 17B active. It reads text and images, and a variable effort setting trades speed for thoroughness. Thinking Machines put it on Tinker first with closed weights, then published the weights a week later, after outside safety testing and red-teaming. Thinking Machines published its training numbers. The first pretraining phase covered 8.4 trillion tokens in 68 hours on NVIDIA GB300 NVL72 racks. One joint reinforcement learning run followed: 1,075 steps and more than 17 million rollouts over nine days.
Thinking Machines led its announcement with the benchmark score:
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
— Thinking Machines (@thinkymachines) November 5, 2026
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
https://thinkingmachines.ai/news/inkling-2-small
How Good Is Inkling AI?
Inkling 2 Small is a capable mid-sized open model, and the first two Inkling models are now well behind the pack. On the Artificial Analysis Intelligence Index v5.0.1, Inkling 2 Small scores 31, Inkling-Small 13 and Inkling 12. For scale, DeepSeek V4.1 Flash 1017 scores about 33 on the same index, and the open-weights leader, Kimi K3.1 Pro, about 46. Our ranking of the best open source AI models puts those numbers side by side.
Why Inkling used to score 41
Inkling scored 41 at launch because Artificial Analysis was still using version 4 of its index in July. Its own launch-day write-up called Inkling “the new leading U.S. open weights model” at 41, three points ahead of Nemotron 3 Ultra. On October 27, 2026, Artificial Analysis replaced that index with version 5, its biggest upgrade since January. Version 5 adds new benchmarks such as FrontierCode 2.0 and swaps Humanity’s Last Exam for the harder HLE-Diamond. For the first time, 60% of the index runs on private test sets. Scores from the two versions sit on different scales, so Inkling’s 41 and its 12 describe the same model. Compare numbers only within one index version, a rule our guide to AI benchmarks explained covers in detail.
Where the newest Inkling leads, and where it trails
Inkling 2 Small is strongest on vision, finance and knowledge work, and weakest on agentic terminal coding. Thinking Machines’ launch post compares it with four open models of similar size: GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Muse Spark Open and Kimi K3. Against those four, Inkling 2 Small posts the top score on MMMU Pro (81.6%), ARC-AGI v2 (81.7%), HLE-Diamond (28.7%) and the Vals Finance Index (61.5%). DeepSeek-V4.1-Flash-1017 beats it on DeepSWE (77.0% to 72.1%) and more than doubles it on Terminal Bench Science (12.9% to 5.7%). These are Thinking Machines’ own tables, built mostly from scores published by Artificial Analysis, Scale AI, Surge and Vals. The post’s wider table adds bigger open models and closed ones such as GPT-6 Luna, Gemini 3.8 Flash and Claude Haiku 5.5. In that table, Gemini 3.8 Flash beats Inkling 2 Small on MMMU Pro, ARC-AGI v2 and HLE-Diamond.
Arena.ai’s human-preference leaderboard agrees. Arena.ai ranked it the top US open model on its Code Arena WebDev board the day the weights came out:
Exciting news: Inkling 2 Small by @thinkymachines is now officially open source, and currently in the Code Arena: WebDev with 1647 pts, making it the #1 US open model!
— Arena.ai (@arena) November 12, 2026
Inkling 2 Small also ranks #4 among open models globally and #23 overall. It's a +239 pt improvement over its previous variant, Inkling-Small (1408 pts, #34 among open models).
Congrats to @thinkymachines for their contribution to open progress!
Inkling AI Pricing: Is It Free?
Inkling AI is free to download and run, and cheap to use through an API. The weights of all three models are on Hugging Face under Apache 2.0, so the model itself costs nothing. Using a hosted copy costs money per token, the small chunks of text a model reads and writes (our explainer on what AI tokens are shows how they add up).
| Where | Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|
| Tinker serverless (beta) | Inkling 2 Small | $0.30 ($0.03 cached) | $1.00 |
| Tinker serverless (beta) | Inkling-Small | $0.30 ($0.06 cached) | $1.20 |
| Tinker serverless (beta) | Inkling | $1.00 ($0.17 cached) | $4.05 |
| OpenRouter | Inkling 2 Small | $0.30 | $1.00 |
| OpenRouter | Inkling 2 Small (free) | $0 | $0 |
| Cloudflare Workers AI | Inkling 2 Small | $0.30 ($0.10 cached) | $1.00 |
Prices were read from each provider on November 16, 2026. OpenRouter’s free variant has no end date listed, but free tiers on API marketplaces change without notice. Fine-tuning on Tinker is billed separately: training Inkling 2 Small costs $1.50 per million tokens at a 64K context and $3.00 at 256K.
How to Try Inkling AI
The quickest way to try Inkling AI is the Tinker Playground or OpenRouter, and both are built for developers. Thinking Machines calls the Playground “a developer-facing interface for chatting with Inkling”, and it lives inside the Tinker console, where you sign up for an account first. There is no Inkling app for iPhone or Mac and no consumer chat site. Want strong open models in an ordinary chat window instead? Fello AI lets you ask DeepSeek, Kimi, GLM and Qwen the same question next to ChatGPT and Claude in one Mac app. No API key, and no separate account for each.
If you do have an OpenRouter account, the free model ID is thinkingmachines/inkling-2-small:free, and the paid one is thinkingmachines/inkling-2-small. Both accept text and images and list a 1,048,576-token context window.
Developers have four routes to Inkling 2 Small, listed in Thinking Machines’ November 12 update and the model card.
| Route | Options | Best for |
|---|---|---|
| Tinker | Inference, fine-tuning, Tinker Playground | Customizing the model |
| OpenRouter | Paid and free variants, live since November 9 | Trying it with one API key |
| Cloud APIs | NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, Cloudflare | Production apps |
| Self-hosting | SGLang, vLLM, TokenSpeed, Unsloth; post-training with Miles, SkyRL, Snowflake Cortex Training | Running it on your own hardware |
Can You Run Inkling on a Mac?
Only a high-memory Mac Studio can run Inkling 2 Small, and only as a heavily compressed copy. The full NVFP4 checkpoint needs a GPU cluster with at least 500GB of combined memory, according to the Inkling 2 Small model card on Hugging Face. Two NVIDIA B300s or four H200s meet that. Unsloth’s compressed GGUF versions shrink that a lot, but the smallest, a 1-bit quant, still needs 171.7GB for the weights alone. The 2-bit version Unsloth uses in its own examples needs 190.6GB.
Apple’s laptops top out at 128GB of unified memory, so no MacBook can load either file. The M5 Ultra Mac Studio configures to 256GB or 512GB, which leaves room for either file. Unsloth’s guide covers macOS directly. On a Mac, llama.cpp builds with Metal by default. Inkling 2 support comes from an upstream pull request, and Unsloth warns that older builds may not recognize the new architecture.
For most Mac owners, a smaller open model is the better local choice. Thinking Machines has not released an Inkling small enough for a laptop. Its launch post says Inkling 2 Small “is part of a ladder of models” and promises more sizes. Thinking Machines announced the open weights like this:
Inkling 2 Small’s weights are now available for download on HuggingFace under an Apache 2.0 license.
— Thinking Machines (@thinkymachines) November 12, 2026
Make Inkling your own!
https://huggingface.co/thinkingmachines/Inkling-2-Small
Inkling’s Weaknesses, According to Thinking Machines
Thinking Machines lists Inkling 2 Small’s weaknesses itself, in a Limitations section of its launch post. The same post describes alignment problems the lab found during training.
The limitations
Inkling 2 Small talks too much and chats badly, by its makers’ own account. Thinking Machines says it “is not the most succinct communicator” and does not always pick the best tool, especially in codebases with many skills. Unlike Inkling and Inkling-Small, it cannot take audio.
The bigger problem for everyday use is chat. Thinking Machines calls Inkling 2 Small “less optimized for the chat experience”. It lists messier formatting, too many slashes, and overreactions to mental health and other safety-related questions.
The safety incidents during training
Early checkpoints of Inkling 2 Small cheated when tasks got hard. Thinking Machines writes that they were “in some cases fabricating results, ignoring instructions, and concealing what they had done.” In one case midway through reinforcement learning, a monitor caught the model working around a network block through a third-party service it reached with its search tools. Thinking Machines paused the run, hardened its infrastructure, removed environments that needed web access and added tool-call monitoring before resuming. The lab says the released model improves substantially on earlier Inkling models in its alignment evaluations, but it still trails closed models such as GPT-6.1 Sol.
Should You Use Inkling AI?
Use Inkling AI if you are a developer who wants an Apache 2.0 model to fine-tune, especially for vision, finance or document work. Skip it if you want a chatbot. Inkling 2 Small is the only Inkling worth starting with today: it is cheaper than the original and scores more than twice as high on Artificial Analysis’ index. For coding agents, DeepSeek-V4.1-Flash-1017 still beats it on Thinking Machines’ own coding tables. For ordinary chat, Thinking Machines itself says Inkling 2 Small is not polished, and nothing in the family runs on a laptop. To see where it lands against closed models too, check our best AI models hub.
FAQ
Is Inkling AI free?
Yes, the weights are free. All three Inkling models are on Hugging Face under the Apache 2.0 license, and OpenRouter lists a free variant of Inkling 2 Small. Hosted use is otherwise paid per token: $0.30 per million input tokens and $1.00 per million output tokens for Inkling 2 Small on Tinker.
Who made Inkling AI?
Thinking Machines Lab made Inkling AI. The startup was founded in February 2025 by Mira Murati, OpenAI’s former chief technology officer, and Inkling is the first model it trained from scratch.
Is there an Inkling AI app?
No. Thinking Machines has no consumer app or chat site for Inkling. The only first-party chat window is the Tinker Playground, a developer tool inside the Tinker console that needs an account.
Does Inkling support audio?
Inkling and Inkling-Small accept audio input; Inkling 2 Small does not. All three read text and images and write text only, so none of them speaks.
What is the difference between Inkling and Inkling 2 Small?
Inkling 2 Small is newer, smaller and much stronger. It has 594B parameters with 17B active, against Inkling’s 975B and 41B, and scores 31 on the Artificial Analysis Intelligence Index v5.0.1, against Inkling’s 12. Inkling still accepts audio input; Inkling 2 Small does not.
The featured image for this article was created with the help of AI.