They Looped the Model's Middle Layers Three Times: Inkling 2 Mini, 26B Parameters Running on a Laptop
Thinking Machines releases Inkling 2 Mini: a 26.4B MoE looped transformer whose middle layers are looped three times; it can run locally with about 20GB of memory. DeepSWE v1.1 reaches 48.2%, ranking first among open sub-100B models. Unsloth has already released a GGUF quantized version that can run on 16GB RAM.
Thinking Machines released Inkling 2 Mini today. It is a 26.4B total-parameter MoE model with 3.4B active parameters and about 20GB memory footprint, and its goal is clear: let you run an agentic model with enough reasoning capability on a laptop.
Its most distinctive aspect is the architecture. Inkling 2 Mini is a looped transformer. In a standard transformer, each token passes through every layer once in sequence. Inkling 2 Mini's 24 layers are divided into four blocks, and the middle two blocks (layers 7-18) repeatedly compute each token three times. That is, each token actually goes through 48 layers of computation, but only 24 layers of weights are stored.
The benefit of doing this: the model can perform deeper internal reasoning before outputting a token, so the chain-of-thought it needs to write out is shorter, especially for tasks that depend on reasoning quality, such as agentic coding. Official numbers are DeepSWE v1.1 48.2%, Vals Index v2.2 46.5%, and AA Index 20.1 (tied for highest)-all in the top tier among open-weight models with under 100B total parameters. For comparison, Qwen3.8-27B scores 42.2% on DeepSWE and 73.8% on SWE-bench Pro v2 Hard; Inkling 2 Mini scores 77.9%.
But loops improve reasoning, not knowledge. On factual benchmarks such as AA-Omniscience, it is on par with other 26B models. Its solution is to use tools: BrowseComp 79.0%, DeepSearchQA 85.3%, and MCP Atlas 82.7%, all the highest among open models of the same size. Put simply, it knows what it does not know, then goes to look it up.
Training is also interesting. The model was natively trained at NVFP4 4-bit precision: 20 layers NVFP4 + 4 layers MXFP8, and deployment uses the same precision as training, with no second quantization loss. Pruning was also done along the way: during training, two-thirds of the experts were pruned, then healed through distillation with 1T tokens. So the final model has only 26.4B total parameters but retains the performance of a larger internal model.
Local deployment is the key point. Officially, it can run on a Mac with 32GB RAM, and DGX Spark works too. Unsloth moved quickly and has already released a Dynamic GGUF quantized version, UD-Q4_K_XL 17.6GB, which can run on 16GB RAM, with a 98.0% top-token agreement rate with the 8-bit version. SGLang also released a support video.
Of course, actually running it locally is another matter. Some users tried it on a 3080 10GB + 32GB RAM setup and got only 2.89 t/s at 4096 context-barely runnable. Others said outright that the architecture is wild, and that there is finally a model you can actually download and play with.
There is a deeper point here. The loop architecture lets the model do more computation internally, so does its chain-of-thought become harder to monitor? Thinking Machines specifically ran experiments on this and also released a variable-loop research model.
The conclusion is somewhat subtle: more loops do improve the model's no-CoT capability-for example, the Think Fast time span rises from 16 seconds with 1 pass to 37 seconds with 3 passes. But the chain-of-thought does not become unreadable-it gets shorter, from 3703 tokens to 1661 tokens; sentence length and vocabulary distribution are unchanged, and it is just more direct and does less backtracking to fix errors. More importantly, the model has very little control over its own chain-of-thought. On CoT-Control-QA, for CoTs of 750-1250 tokens, the probability of following instructions is no more than 0.1%. In other words, so far there is no sign that looping makes monitoring harder.
Inkling 2 Mini is released under Apache 2.0, with weights on Hugging Face, and Tinker supports inference and fine-tuning. For a 26B model, it pushes reasoning efficiency to a new position. Next, we'll see how the community uses it.
Official blog: https://thinkingmachines.ai/news/inkling-2-mini/
Hugging Face: https://huggingface.co/thinkingmachines/Inkling-2-Mini
Unsloth GGUF: https://huggingface.co/unsloth/Inkling-2-Mini-GGUF
发布时间: 2026-11-21 07:47