Key takeaways
- Thinking Machines Lab released Inkling 2 Mini on 20 November 2026 as an open-weight model, with weights available on Tinker and Hugging Face.
- It is a 26.4-billion-parameter looped Mixture-of-Experts transformer, with 3.4 billion active parameters per token and a memory footprint of about 20GB.
- The model runs on a Mac with 32GB of memory or a single NVIDIA DGX Spark, bringing serious coding AI to ordinary hardware.
- It scores 79.2% on SWE-bench Verified, 48.2% on DeepSWE v1.1 and 46.5% on the Vals Index v2.2, according to the lab.
- It is a new, smaller model, not the 594-billion-parameter Inkling 2 Small released earlier in November.
What is Inkling 2 Mini?
Inkling 2 Mini is the newest model from Thinking Machines Lab, the AI company founded by former OpenAI chief technology officer Mira Murati. It arrived on 20 November 2026, just eight days after the lab released its much larger sibling, Inkling 2 Small. Where that model has 594 billion parameters and needs serious server hardware, Mini is designed to fit on the machine on your desk.
Thinking Machines Lab says this looping lets the model apply more reasoning before it produces its next token, which matters most for agentic coding: long, multi-step jobs where an AI writes code, runs tests, reads the errors and tries again. The company describes the idea as reasoning longer without writing its reasoning down, which also matters for monitoring what the model is doing.
The numbers that matter
Inkling 2 Mini is a Mixture-of-Experts model, a design in which only some of the model's experts handle each token. It has 26.4 billion total parameters, with 3.4 billion active per token. Because the looped layers are visited three times, the effective figure is about 5.4 billion weight activations per token. The whole model fits in roughly 20GB of memory, thanks to weights trained and deployed mostly in NVFP4 precision, with the last four layers in MXFP8.
That small footprint is the point. Thinking Machines Lab says the model can run on a Mac with 32GB of unified memory, or on one NVIDIA DGX Spark desktop AI computer. The lab reports that researchers ran it offline on a MacBook Pro and had it build a playable flight simulator in a few minutes. The model also accepts text and image input, supports a context window of up to one million tokens, and is released under the Apache 2.0 licence, which allows broad commercial use.
The weights are available on Hugging Face, and the model is available on Tinker, the lab's own platform, for inference and fine-tuning. A GGUF build is also available for tools such as Ollama and llama.cpp, which is how many people run models locally.
How it performs
On the lab's own benchmark table, Inkling 2 Mini leads every open-weights model with fewer than 100 billion total parameters on several coding tests. It scores 79.2% on SWE-bench Verified, a widely used test of whether an AI can fix real software issues from GitHub. It scores 48.2% on DeepSWE v1.1, ahead of Qwen3.8-27B at 42.2%, and 46.5% on the Vals Index v2.2. Its SWE-bench Pro v2 Hard score is 77.9%, ahead of Qwen3.8-27B at 73.8%.
Independent evaluator Artificial Analysis gave the model a score of 20 on its Intelligence Index, matching Qwen3.8 27B while using far fewer active parameters. Artificial Analysis also found that Inkling 2 Mini uses about 38% more output tokens than Qwen3.8 27B for the same score, but about 2.8 times less compute, and costs roughly six times less per task.
These are important numbers, but they come with a caveat that applies to every model launch. Most of the headline scores are company-reported results, and different testing setups can change the outcome. The model is not ahead of every rival on every test: Thinking Machines Lab itself notes that it sits behind Qwen3.8-27B on some benchmarks. Readers should treat the scores as a strong first signal, not as a final verdict.
Why a small model matters
The most interesting thing about Inkling 2 Mini is not a benchmark score but where it can run. The best AI models have mostly lived in data centres, reached through an API, with your code and data travelling to someone else's servers. A model that runs on a laptop changes that. Developers can work offline, keep private code on their own machine, and avoid paying per-token API bills for every coding session.
Limitations to keep in mind
Small models have real limits. A 26-billion-parameter model, however clever its architecture, does not know as much as a model ten or twenty times its size. The lab itself frames the loop as buying reasoning depth, not knowledge. For tasks that depend on broad factual knowledge, a larger model will usually do better.
Local running also has practical demands. The model needs a Mac with at least 32GB of memory, or equivalent hardware, which rules out many ordinary laptops. Running it well also takes some technical setup, through Ollama, llama.cpp or Tinker, rather than simply opening a website. And while the weights are open under Apache 2.0, the training data is not fully open: the lab says the model trained on one trillion tokens distilled from a larger internal Inkling model, which means its knowledge ultimately comes from a bigger system that is not public.
There is also the question of how it behaves on tasks beyond coding. The lab's strengths are concentrated in agentic coding and tool use, and the model is behind some rivals on other benchmarks. Anyone planning to use it as a general-purpose assistant should test it on their own work first.
What happens next
The release became the most popular model on Hugging Face on its launch day, which suggests strong interest from developers. The next test is practical: how well it performs in real coding workflows, in tools such as Claude Code-style agents and local editors, over days and weeks rather than in a benchmark table.
Thinking Machines Lab has now released two open-weights models in November 2026, and the pattern is clear. The lab is betting that open models, running on hardware people already own, are a serious alternative to closed API models. Inkling 2 Mini is the strongest evidence yet for that bet, and the coming months will show whether developers agree.
Conclusion
Inkling 2 Mini is a small model with a big idea. By looping its middle layers three times per token, Thinking Machines Lab has fitted strong agentic coding ability into about 20GB, a model that runs on a 32GB Mac or a single DGX Spark. Its 79.2% score on SWE-bench Verified and its open Apache 2.0 licence make it one of the most interesting releases of the month. It will not replace large models for every job, and its headline scores are mostly the lab's own, but for developers who want a capable coding model running on their own machine, it is a genuine new option.
Frequently asked questions
Can Inkling 2 Mini really run on a Mac?
Yes, according to Thinking Machines Lab. The model has a memory footprint of about 20GB and is designed to run on a Mac with 32GB of unified memory, or on a single NVIDIA DGX Spark. It is also available through Ollama and llama.cpp builds for local use.
Is Inkling 2 Mini free to use?
The weights are released under the Apache 2.0 licence, which allows broad use including commercial use. The model is available on Hugging Face and on Tinker, the lab's platform, for inference and fine-tuning. Running it locally costs nothing beyond the hardware.
How is Inkling 2 Mini different from Inkling 2 Small?
Inkling 2 Small is the lab's 594-billion-parameter model, released on 12 November 2026, and it needs server hardware. Inkling 2 Mini is a new, much smaller model with 26.4 billion parameters, designed to run on a laptop or desktop AI computer, using a looped transformer design that reuses its middle layers three times per token.
Researched with AI assistance and reviewed for accuracy.
Follow on WhatsApp
Follow Daily Tech & AI on WhatsApp for the latest tech news, AI updates and research breakthroughs: https://www.whatsapp.com/channel/0029Vb9XzdHDJ6H4TFNAl61r
Comments
Post a Comment