Inkling 2 Mini: Thinking Machines Opens a 26B Looped Model for 32GB Macs
Inkling 2 Mini is a 26.4B-parameter open-weight model from Thinking Machines Lab, released on 20 November 2026 under Apache 2.0. It activates 3.4B parameters per token and loops its middle 12 layers three times. The single checkpoint is about 20GB, and the lab says it runs offline on Apple silicon Macs with at least 32GB of memory.
On this page
What Thinking Machines released on 20 November
Thinking Machines Lab, the company behind the much larger Inkling model, has published Inkling 2 Mini, a smaller model it describes as fast, strong at agentic work, and small enough to run on a laptop. The weights are on Hugging Face under the Apache 2.0 license, which allows commercial use and modification. The model takes images as well as text, because a vision encoder is built in, and the model card lists a context window of up to 1 million tokens.
Alongside it, the lab released a research variant trained to run a variable number of loops, which it used for experiments on how the model reasons. The model is also available for chat and fine-tuning through Tinker, the company's own platform, and through third-party inference providers.
How looped layers and experts fit together
Inkling 2 Mini combines two ideas for getting more out of fewer weights. The first is a sparse mixture of experts: the 24-layer model holds 26.4B parameters in total, but each token is routed to only 8 of 128 experts per layer, so about 3.4B parameters do the work for any one token. The lab says it pruned the model from 384 experts per layer, keeping the 128 used most.
The second is looping. The middle 12 layers run three times with the same weights, so each token passes through 48 layer steps while only 24 layers are stored. The model card counts this as about 5.4B effective active parameters per token. As with other looped designs, the cost is extra compute per token instead of extra memory, which is the right trade for machines with limited RAM. CuriousLM covered a smaller version of the same idea in SenseNova-Loop-9B earlier this week.
Coding and agent scores, with caveats
The lab's headline claims are on software engineering. It reports 48.2% on DeepSWE v1.1 and 77.9% on the hard split of SWE-bench Pro v2, against 42.2% and 73.8% for Qwen3.8-27B, a dense model of similar size. It also reports 79.2% on SWE-bench Verified and 79.0% on BrowseComp.
The picture is not one-sided. On the lab's own tables Qwen3.8-27B leads on Terminal Bench v2.1, ARC-AGI-1, MultiChallenge and ComplexConstraints, and the two tie on the Artificial Analysis index. The blog is also frank that factual knowledge depends mostly on total parameter count, so a 26B model will hallucinate more than very large ones. Almost all of these numbers come from the developer. TestingCatalog relays an Artificial Analysis figure of roughly $0.59 per index task against $3.53 for Qwen3.8-27B, but that uses Thinking Machines' own pricing, not a published price list.
What you need to run it at home
For local use the key detail is the checkpoint. The lab ships one file set of about 20GB, with most layers in NVFP4 and the last four in MXFP8, and says these are the precisions the model was trained in, so there is no conversion loss from quantizing afterwards.
The blog says any Apple silicon Mac with at least 32GB of RAM can run it through Ollama, and shows it running offline on a 48GB MacBook Pro. The model card adds a single DGX Spark or GPUs with 32GB or more of VRAM as recommended options. Quantized GGUF builds are available through Unsloth and llama.cpp, and serving engines such as vLLM and SGLang are supported. The Hugging Face page already lists dozens of community quantizations.
That puts a strong open coding agent within reach of a high-end laptop with no cloud account and no data leaving the machine. Treat the benchmark lead as provisional until independent tests arrive, and check the memory headroom on your own hardware, because long contexts need far more than the 20GB of weights.