A large model that fits on consumer hardware
Thinking Machines Lab has released Inkling 2 Mini, an open-weight language model with 26.4 billion total parameters and 3.4 billion active parameters, which according to the source reaches 5.4 billion effective per token thanks to the looping mechanism. The main novelty is not the size, but the way the model exploits its weights: instead of adding parameters, it reuses the middle layers three times for every token. The result is a checkpoint of about 20GB, compatible with a Mac with at least 32GB of memory or with a single NVIDIA DGX Spark.
The model card indicates a context window of up to one million tokens and the ability to accept text and images while returning only text. Distribution is under the Apache 2.0 license on Hugging Face, with hosted inference and fine-tuning via Tinker. Quantized GGUF builds are available via Unsloth and llama.cpp, and serving support covers Transformers, vLLM, SGLang, Ollama, NVIDIA Dynamo, and TokenSpeed. The source does not report Tinker's price for this version.
How looping works and why it was chosen
The model stores 24 layers, organized into four blocks of six. The first and last blocks are run only once, while the two middle blocks are repeated three times with shared weights: this yields 48 layer applications per token starting from 24 layers of weights. Every pass after the first receives a message from the previous pass, conveyed by Attention Residuals and a learned gate. The feed-forward layers use 128 experts, of which 8 are routed per token.
According to Thinking Machines, looping the middle layers produces an equivalent compute gain of between 1.51x and 1.91x, while repeating the final layers stops at between 1.24x and 1.36x; the middle-loop variant would have won 27 out of 28 evaluations. The mechanism, however, shifts the cost: a baseline without looping used 7.5% less training compute, but required 140% more inference compute to match the loss of the looping model on 512k-token agentic traces. It is a design choice that prioritizes efficiency at use time over training efficiency.
Reported results and acknowledged limitations
All Inkling scores cited by the source are self-reported by Thinking Machines and should be treated as such. The strongest area is coding: 79.2% on SWE-bench Verified in bash-only mode and 77.9% on SWE-bench Pro v2 Hard; on DeepSWE v1.1 it scores 48.2%, versus the 42.2% that Qwen's model card attributes to Qwen3.8-27B. On Terminal Bench v2.1 it stops at 71.6%, behind Qwen3.8-27B's 73.0%. Agentic search records 79.0% on BrowseComp with compaction and 82.7% on MCP Atlas; for documents and vision, the model reports 92.8% on OmniDocBench v1.6 and 74.9% on MMMU Pro.
The weak points concern above all knowledge and dialogue. The model scores -8.6 on AA-Omniscience with a non-hallucination rate of 63.4%, and the source observes that looping did not help knowledge benchmarks, which depend mainly on the total number of parameters. On MultiChallenge it gets 50.0% and on ComplexConstraints v1.1 33.4%, with lower results than larger models also on HLE-Diamond (15.2%) and CritPt v2. The company itself states that it is not the best choice for multi-turn chat or prompts with many constraints.
On energy efficiency, Thinking Machines reports 0.703 on Stanford's Intelligence per Watt metric on DGX Spark, versus 0.430 for Qwen3.6-35B-A3B and 0.159 for Gemma-4-26B-A4B. The source warns, however, that each column uses the respective manufacturer's model card and that harnesses and settings differ: cross-vendor comparisons should be read as indicative, not as homogeneous measurements. The self-reported scores and the price of hosted inference also remain to be independently verified.
