Thinking Machines releases Inkling 2 Mini: 26B parameters and 20.1 points on the AA Index
Thinking Machines has shrunk Inkling 2 Small from 594B to 26B parameters, and the new Inkling 2 Mini runs on a laptop. It is a looped transformer: the same layers are run several times instead of building a wider network. On the Artificial Analysis agentic index, the model scored 20.1 points, compared with 31 for the larger version. For local agents, it is a workable option if quality below the flagship is acceptable.
AI-processed from X: @stretchcloud; edited by Hamidun News
The main complaint about open-weight models in recent months sounds the same: they may be strong, but few people can run them themselves. Thinking Machines' Inkling 2 Small, with 594B parameters, drew exactly this reaction. Two weeks later, the lab responded with Inkling 2 Mini, a 26B-parameter model. According to the company, it fits on a laptop and is suited to agentic tasks. On the Artificial Analysis index for agentic work, Mini scores 20.1 points. The larger Small scored 31 on the same index.
How to fit reasoning into 26B parameters
Inkling 2 Mini is built as a looped transformer. In a conventional model, each layer is separate: the more layers, the more unique weights and the heavier the file. Here, the same set of layers is applied several times in a row. Imagine one editor who rereads a draft ten times instead of ten different editors, each editing it once. There are fewer unique weights, but more computation per token. The company trades storage size for compute time.
For practitioners, this means the model takes up little space and fits in laptop memory, but each response requires a noticeable amount of computation. This suits local deployment, where fitting the hardware matters more than maximum speed.
Where the open model market is heading
The author of the analysis behind this news places Inkling in a broader picture. DeepSeek V4.1 Flash has 552B parameters in total but activates about 16B per token. Qwen from Alibaba and GLM from Zhipu, according to the author, have similar sparse Flash versions. The idea is the same: a huge total parameter count with only a few percent active at any moment. This lets a consumer GPU or a powerful laptop handle what once required server racks.
Labs are thus pursuing cheap local deployment by two different routes. Some make the network sparse, so part of the weights sleep at any given moment. Others, like Thinking Machines, make it looped, so there is one set of weights and several passes. The overall conclusion is not about growing size, but about how to get reasoning without expensive infrastructure.
Transparency as part of the product
The release scheme is also unusual. Before the general release, Thinking Machines gave early access to Goodfire and NDIF, groups working on interpretability. They were to study how information moves between loops inside the model. Mira Murati's lab, which reportedly raised $2B at a $10–12B valuation, is betting on a research-driven approach and positioning itself against OpenAI and Anthropic.
What to consider
First, 20.1 points is noticeably less than Small's 31. This is not a replacement for the flagship but a compromise: you gain locality and pay with quality. For simple agentic scenarios where data cannot be sent to the cloud, this may be enough, but you should test on your own tasks, not just rely on the index.
Second, looped architecture has a downside that the author warns about directly. The weight file is small, but the model's internal state is harder to inspect than in a standard transformer: the same layers operate at different steps, and it is not easy to understand what happens on each pass. Small size on disk does not mean simplicity. That is why early access for Goodfire and NDIF looks less like a marketing detail and more like a response to a real auditing problem.
Finally, keep the source in mind. This is an X post that retells and comments on the news, and some of the claims about DeepSeek, Qwen and GLM belong to the post's author, not to Thinking Machines. The figures of 26B and 20.1 points come from the lab's announcement. There are not yet any independent measurements of speed or quality on a real laptop.
Practical takeaway
If you are building local agents or want to keep a model inside your own environment, Inkling 2 Mini is worth adding to your test list alongside the sparse Flash models. Compare on three parameters: quality on your tasks, response time on your hardware, and the ability to verify the model's behavior. The last point matters especially in regulated industries, where auditability is as important as accuracy.
- Nov 21, 2026Anthropic mathematician says Claude helped prove the Riemann hypothesis in six weeks
- Nov 21, 2026Startup Osmo Trains AI on Smells Using 10 Million Human Ratings and Generates Fragrances From Text
- Nov 21, 2026North Korea Shows First Chatbot "Source of Wisdom": Science Only, No Politics
- Nov 21, 2026OpenAI Opens ChatGPT Spaces to Plus Subscribers: Shared Documents with AI
Key Questions Answered
How to fit reasoning into 26B parameters?
Inkling 2 Mini is built as a looped transformer. In a conventional model, each layer is separate: the more layers, the more unique weights and the heavier the file. Here, the same set of layers is applied several times in a row. Imagine one editor who rereads a draft ten times instead of ten different editors, each editing it once. There are fewer unique weights, but more computation per token. The company trades storage size for compute time.
Where the open model market is heading?
The author of the analysis behind this news places Inkling in a broader picture. DeepSeek V4.1 Flash has 552B parameters in total but activates about 16B per token. Qwen from Alibaba and GLM from Zhipu, according to the author, have similar sparse Flash versions. The idea is the same: a huge total parameter count with only a few percent active at any moment. This lets a consumer GPU or a powerful laptop handle what once required server racks. Labs are thus pursuing cheap local deployment by two different routes.
What are the key details on Transparency as part of the product?
The release scheme is also unusual. Before the general release, Thinking Machines gave early access to Goodfire and NDIF, groups working on interpretability. They were to study how information moves between loops inside the model. Mira Murati's lab, which reportedly raised $2B at a $10–12B valuation, is betting on a research-driven approach and positioning itself against OpenAI and Anthropic.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.