Table of Contents
On November 5, 2026, Thinking Machines Lab released Inkling 2 Small, its second-generation AI model, and the launch quietly became one of the most interesting releases of the month. The model is built for serious coding work, not casual chat: it writes code, fixes bugs, and handles long autonomous software tasks with very little hand-holding. If you follow AI news, this model is worth your attention, because it shows where the newest labs are heading — smaller active models, sharper coding skills, and open weights on the way. Here are seven stunning facts you should know about it.
1. Inkling 2 Small Packs 594 Billion Parameters
Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters, of which only 17 billion are active for any given token. That design is the whole point. Instead of firing up the entire model for every word, it routes each task to a small team of specialist experts, which keeps the model fast and affordable to run while preserving the knowledge of a much larger system. For developers, that means frontier-level coding ability without the heavy compute bill that usually comes with it. Inkling 2 Small proves that smart routing can matter more than raw size.
2. Inkling 2 Small Is Built for Coding and Autonomous Software Work
Unlike chat-first models that try to do everything, this model has a clear job: coding and autonomous software engineering. Thinking Machines Lab trained it for the kind of work developers actually do — writing new features, debugging broken code, refactoring old projects, and running multi-step tasks on its own. The company reports a score of 86.3% on SWE-bench Pro v2 (Hard), a demanding benchmark that tests real software fixes, and 72.1% on DeepSWE v1.1. Those are company-reported figures, but they signal a model aimed squarely at professional developers rather than casual users.
3. It Understands Images Too
The model is multimodal, which means it takes image input as well as text. That matters more than it sounds. A coding model that can see a screenshot of a broken layout, a chart, or a whiteboard sketch can help in ways a text-only model cannot. You can show it an error screen, a design mockup, or a diagram of your system, and it can work from what it sees. Combined with its adjustable effort setting, which lets you trade speed for deeper reasoning, it adapts to quick questions and hard problems alike.
4. You Can Use It on Tinker Right Now
The model is available today through Tinker, Thinking Machines Lab’s own training and inference platform. Developers can try it in the Tinker Playground, use it for sampling, and even fine-tune it for their own projects. That is a big deal for teams that want to shape a model around their own codebase instead of accepting a one-size-fits-all assistant. Tinker is the company’s home base for the model, and it is where the first users are putting it through its paces on real software work.
5. Inkling 2 Small’s Weights Are Closed at Launch — But Open Weights Are Coming
Here is the twist: the downloadable weights are closed at launch. Thinking Machines Lab says it will release them under the Apache 2.0 license the following week, once external safety testing wraps up. That is a deliberate, careful rollout — test first, open later — and it matches how the lab handled its earlier Inkling models, which are already available as open weights. If the plan holds, developers will soon be able to download, run, and build on Inkling 2 Small commercially, with no licensing fees. For anyone who likes open models, this is the part to watch.
6. A Nine-Day Training Run Powered the Model
Behind the launch is a remarkable training effort. Thinking Machines Lab reports a nine-day reinforcement-learning run involving more than 17 million rollouts across 2,048 GPUs. Reinforcement learning is the technique that teaches a model by rewarding good answers and correcting bad ones, and at this scale it is how a model learns to handle long, messy, real-world coding tasks instead of just short textbook problems. The scale of the run explains why the model feels so capable on autonomous work: it practised millions of software tasks before it ever met a user.
7. Inkling 2 Small Scores 31 on the Artificial Analysis Intelligence Index
Thinking Machines Lab reports a score of 31 on the Artificial Analysis Intelligence Index version 5.0, a widely watched benchmark of overall model ability. That puts the model well above its first-generation predecessor and in serious company among recent releases. Benchmark scores never tell the whole story — real-world coding tests matter more — but a strong index score is a useful first signal. Combined with its coding benchmarks, it suggests a genuine step forward, not just a bigger number on a spec sheet.
How It Fits the November Model Rush
November 2026 has been an extraordinary month for AI releases. Just a day before it arrived, Tencent Hy4 launched with vision and computer use, and earlier in the week Gemini 4 Carbon took the top spot on the Artificial Analysis Intelligence Index. Inkling 2 Small fits this wave perfectly: it is another large model, but one with a narrower, sharper focus on coding.
Where other releases chase general intelligence, Thinking Machines Lab is betting that developers want a specialist. If you create images with AI as well as code, our Midjourney AI review shows how a specialist image model compares. The lesson of the month is clear: the best model for your work may no longer be the biggest one, but the one built for your job.
Should You Try It?
If you write code for a living, yes — try it in the Tinker Playground and see how it handles your real tasks. If you are waiting for open weights, the Apache 2.0 release planned for next week is your moment: a 594-billion-parameter coding model you can download and fine-tune is a rare gift. Either way, Inkling 2 Small is a release worth watching, because it points to a future where powerful models are smaller, cheaper to run, and open to everyone. You can read the official announcement on the Thinking Machines Lab website for the full details.
Frequently Asked Questions
What is Inkling 2 Small?
Inkling 2 Small is a multimodal AI model from Thinking Machines Lab, released on November 5, 2026. It is a mixture-of-experts model with 594 billion total parameters and 17 billion active per token, built mainly for coding and autonomous software work.
How can I use it?
You can use it now through Tinker, the company’s platform, for playground chats, sampling, and fine-tuning. The downloadable weights are not public yet, but Thinking Machines Lab plans an Apache 2.0 release the week after launch, once external safety testing is complete.
Is it open source?
Not yet. The weights are closed at launch, but the company has promised an Apache 2.0 release next week. Its earlier Inkling models are already available as open weights, which makes the promise credible.
Is it good for coding?
Yes, coding is its main strength. The company reports 86.3% on SWE-bench Pro v2 (Hard) and 72.1% on DeepSWE v1.1, both demanding coding benchmarks. As with any vendor-reported score, the best test is trying it on your own projects.

