26B Model Fits on a Laptop: Inkling 2 Mini Loops Its Middle Layers Three Times
Thinking Machines releases Inkling 2 Mini, with 26.4B total parameters and about 20GB memory, runnable on a laptop. It trades inference for looping middle layers, scores 48.2% on DeepSWE v1.1, and can generate a playable flight simulator offline.
Thinking Machines has released Inkling 2 Mini. The positioning is straightforward: a 26.4B total-parameter MoE with about 20GB memory usage, able to run on a laptop. Mira Murati labeled it on X as a looped transformer, where each token passes through the middle layers three times.
This looping design is the key. The model has 24 layers in total, split into four blocks of 6 layers each. The two middle blocks run three times, while the first and last blocks each run once. The official explanation is that the middle layers are more oriented toward abstract reasoning, while the first and last layers are more responsible for conversion between embeddings and internal representations. Putting the loop in the middle 12 layers yields greater gains across 27 evaluations than looping the last 12 layers.
In terms of parameters, it is an MoE with 26.4B total parameters and 3.4B activated parameters. Counting the loops, the weight activation amount per token is 5.4B. Most weights are trained and deployed with NVFP4, with the last four layers using MXFP8, for total memory of about 20GB. The official statement says a Mac with at least 32GB RAM can run it through Ollama, and it also supports NVIDIA DGX Spark.
On performance, the official numbers are DeepSWE v1.1 48.2%, Vals Index v2.2 46.5%, and AA Index 20.1. Among open-weight models under 100B, it leads on multiple agentic coding benchmarks. SWE-bench Pro v2 Hard 77.9%, BrowseComp 79.0%, DeepSearchQA 85.3%.
Image from X user Md Ismail Šojal's reply; refer to the official model card for exact numbers.
Another change brought by looping is the reasoning process. The model can do more internal serial computation per token, so the written CoT can be shorter. The official team experimented with a variable-loop research model: with three loops, the no-CoT time span rose from 16 seconds with one loop to 37 seconds. On the serial reasoning task N-Hop Lookup, performance continued to improve as the number of loops increased, and there was still improvement at ten loops.
This has a direct relationship to monitoring. Looping makes the model write less reasoning, which in theory could make CoT monitoring harder. At this scale, the official team did not see a clear drop in monitorability. With five loops, average CoT length was less than half that of one loop, but sentence length and the proportion of English words did not change; the model was just more direct and did less backtracking. In agentic sabotage tests, increasing loops also did not noticeably degrade performance.
The official team released two local demos. The first was on a MacBook Pro M5 Pro 48GB in airplane mode, where the model spent a few minutes building a playable flight simulator. The timer in the video is real time, and most of the footage is played at 15x speed.
The second is Thinkyville, a pixel town where 20 residents are driven by the same Inkling 2 Mini on an offline MacBook.
The limitations are also clear. Multi-turn chat and complex constrained instructions are weaker: MultiChallenge 50.0% and ComplexConstraints v1.1 33.4%, both behind Qwen3.8-27B. On reasoning-intensive benchmarks CritPt v2 and HLE-Diamond, it also trails larger models. Looping can increase the number of reasoning steps, but it cannot make the weights remember more facts. On knowledge-intensive tasks it is roughly at the same-size level, and it makes up for shortcomings through search and tool use.
Some users asked in replies whether it can run on an M3 MacBook Pro 16GB, and others asked when it will come to iPhone. The official local threshold is a Mac with at least 32GB RAM, running through Ollama. For quantized versions, look at Unsloth and llama.cpp GGUF. Full weights are available on Hugging Face, and inference and fine-tuning are possible on Tinker.
Small models compressing parameters enough to run on a laptop is no longer new. What is worth watching with Inkling 2 Mini is efficiency: looping middle layers, low-bit training, and healing after pruning—these choices together determine how it spends computation per token. Whether it can stay resident on the desktop ultimately depends not on demos, but on latency, tool calls, and stability in real tasks.
Official release post: https://thinkingmachines.ai/news/inkling-2-mini/
Model card: https://thinkingmachines.ai/model-card/inkling-2-mini/
Hugging Face: https://huggingface.co/thinkingmachines/Inkling-2-Mini
Mira's tweet: https://x.com/miramurati/status/2123872534970208421
发布时间: 2026-11-21 19:05