Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, a small open-weight model that the company says runs on a laptop. The model has 26.4 billion parameters, 3.4 billion of them active per token, and the weights are available on Hugging Face under the Apache 2.0 license. The architecture is what stands out: Inkling 2 Mini is a looped transformer that runs its middle layers three times for every token. Thinking Machines says this delivers more computation per token without storing more weights. All benchmark figures in this article come from the vendor and have not been independently verified.
Inkling 2 Mini at a glance
According to the model card on Hugging Face, Inkling 2 Mini is a multimodal model. It accepts text and images and produces text. The context window reaches up to one million tokens. The target audience, per the vendor, is developers building agentic and tool-use systems, coding assistants, chatbots and retrieval applications.
The key specifications:
- 26.4 billion total parameters, 3.4 billion active per token (mixture of experts).
- Because the middle layers run three times, the vendor counts about 5.4 billion effective active parameters per token.
- 24 layers, with each token routed to 8 of 128 experts in every layer.
- A single checkpoint of about 20 GB in NVFP4 and MXFP8 formats.
- Apache 2.0 license, weights on Hugging Face, plus access through the company’s own Tinker platform.
Alongside the main model, Thinking Machines is releasing a research variant with a variable number of loops. Called TML-d24-varloop-research, it is meant to let researchers compare the effect of different loop counts cleanly.
What a looped transformer does differently
In a standard transformer, every token passes through the layer stack exactly once. Inkling 2 Mini splits its 24 layers into four blocks of six. The first and last blocks run once, while the two middle blocks run three times in a row with the same weights. That gives 48 layer applications per token from 24 stored layers.
The reasoning, as laid out in the Thinking Machines announcement: the middle layers carry more of the abstract reasoning, while the outer layers mostly translate between embeddings and internal representations. In the company’s experiments, looping the middle twelve layers produced larger accuracy gains than looping the last twelve on 27 of 28 evaluations.
A learned gate controls how much of the previous pass’s output flows into the next one. At the start of training it blocks that information entirely, so the model initially behaves like a model without loops. It opens up over the course of training. Thinking Machines describes this as a way to introduce looping in a stable manner.
The trade-off shows up in compute. According to the vendor, a comparable model without loops cost 7.5 percent less compute to train but needed 140 percent more inference compute to reach the same loss on long agentic traces.
4-bit training and pruning
Thinking Machines says Inkling 2 Mini was trained largely in 4-bit precision. The expert layers ran in NVFP4 from pretraining through reinforcement learning, and the last four layers in 8-bit MXFP8. Because the weights ship in the same format they were trained in, the vendor says there is no quality loss from post-training quantization.
The model originally had 384 experts per layer. During training, Thinking Machines removed two thirds of them and kept the 128 most used. The model then trained on one trillion tokens distilled from a larger internal Inkling model to recover its pre-pruning performance. Supervised fine-tuning followed, then a reinforcement learning phase totaling 2.2 trillion tokens over 2,000 steps.
Benchmarks according to the vendor
Thinking Machines positions Inkling 2 Mini as the strongest open-weight model under 100 billion parameters on several coding benchmarks. Figures it cites include:
- DeepSWE v1.1: 48.2 percent
- Vals Index v2.2: 46.5 percent
- SWE-bench Verified (bash only): 79.2 percent
- SWE-bench Pro v2 (Hard): 77.9 percent
- BrowseComp with compaction: 79.0 percent
- MCP Atlas: 82.7 percent
The comparison covers models of similar size such as Gemma 4 31B, Qwen3.8-27B and Nemotron 3.5 Super VL. In the vendor’s own tables, Qwen3.8-27B comes out ahead on some measures, for example Terminal Bench v2.1 (73.0 versus 71.6 percent). Thinking Machines itself notes that evaluation setups differ between models.
The section on weaknesses is unusually candid. For multi-turn chat and instructions with many constraints, the company says Inkling 2 Mini is not the strongest small model: it scores 50.0 percent on MultiChallenge and 33.4 percent on ComplexConstraints v1.1, behind Qwen3.8-27B on both. Looping does not help on knowledge-heavy tests either. The loops improve reasoning but do not let the weights store more facts, because knowledge mostly depends on total parameter count, according to the vendor. The model is meant to close that gap through search and tool use.
Running it locally: hardware and software
For developers, the main draw is that the roughly 20 GB checkpoint can be used locally. Thinking Machines lists the following supported hardware:
- one NVIDIA DGX Spark,
- Apple silicon devices with at least 32 GB of memory, meaning MacBooks, Mac minis and Mac Studios,
- other GPUs and platforms, with at least 32 GB of VRAM recommended.
On the software side, the model card lists Hugging Face Transformers, SGLang, vLLM, TokenSpeed, NVIDIA Dynamo, Unsloth, llama.cpp and the Llama App, plus Ollama and Exo Labs on Apple silicon. According to the Hugging Face blog, GGUF builds in 4 and 8 bits are available for llama.cpp. For reinforcement learning fine-tuning, the vendor names SkyRL, Miles and NVIDIA AutoModel.
As a demonstration, Thinking Machines describes how the model, running offline on a MacBook Pro during a flight, built a playable flight simulator. Demos like this show what is possible but say little about how reliably a model performs day to day. Teams that want to use Inkling 2 Mini for coding agents should test it on their own tasks before it replaces an existing model.
Looping and monitorability
Part of the release concerns AI safety. Many oversight approaches read a model’s chain of thought to spot problematic behavior. That works because a standard transformer can only perform a limited number of computation steps per token, so longer reasoning has to be written out as text. A looped transformer can compute more internally before it outputs anything. That raises the concern that some of its reasoning could escape monitoring.
Thinking Machines studied this with the variable-loop variant. The time horizon the model can handle without a written chain of thought rose from 16 seconds with one pass to 37 seconds with three. The chain of thought got shorter with more loops but, according to the company, stayed legible. Tests for agentic sabotage showed no upward trend as loops were added. One stated goal of the release is for outside researchers to keep investigating these questions.
How it fits in
Inkling 2 Mini joins a wave of open models that recently included Tencent’s Hy4 with open weights and Cloudflare’s Clairic extraction models. The difference is the size class and the architecture: instead of ever more parameters, Thinking Machines relies on repeated layers and native 4-bit training to bring a strong model to consumer hardware.
For teams that cannot or do not want to send data to external APIs, that is a relevant option, especially since the Apache 2.0 license also permits commercial use. Whether the benchmark leads hold up in practice will become clearer as independent tests arrive in the coming weeks. The model is also interesting for research, because it is one of the first freely available looped transformers in this performance class.




