Inkling 2 Mini leads the charts it drew and ties on the one it did not
Inkling 2 Mini is a 26.4B open MoE that runs its middle layers three times per token, claims the sub-100B coding lead and only ties Qwen on the AA Index.

Thinking Machines released Inkling 2 Mini on 20 November, an open-weights model small enough for a laptop. Its pitch is not size. The middle twelve of its 24 layers run three times for every token; Thinking Machines argues that internal depth buys reasoning a 26.4B parameter count would not otherwise reach.
Three passes over one set of weights
Architecturally it is a sparse mixture of experts: 26.4B parameters in total, 3.4B active per token, or 5.4B weight activations counting the repeated reads. The outer six-layer blocks run once; the two middle blocks share weights across three passes, yielding 48 layer applications from 24 stored layers. Thinking Machines reports that looping the middle layers beat looping the last twelve on 27 of 28 evaluations.
The parameter count invites a wrong inference. Weight memory scales with the total; compute per token with the active set. A 26.4B/A3.4B model is not a smaller download than a 26.4B dense one, only cheaper to run.
The table, read carefully
The headline is scoped to its own class: the best scores of any open-weights model under 100B total parameters on DeepSWE v1.1, at 48.2% of tasks resolved, and on Vals Index v2.2, at 46.5%.
The one independent index is flatter. Artificial Analysis's AA Index v5.0 gives Inkling 2 Mini 20.1 and Qwen3.8-27B 20.1; the vendor's own row marks both as tied. Its speed chart is titled "theoretical tokens per second", and the footnote says decoding speed is "a theoretical ceiling at 614 GB/s with BF16 weights and batch size one, including repeated reads across loops". It ends: "Evaluation setups vary by model."
What a 20GB footprint buys
Most weights were trained and are served natively in NVFP4, the last four layers in MXFP8, for about 20GB, with no conversion loss from quantizing. That reaches Macs with 32GB or more via Ollama, Nvidia's DGX Spark, and GGUF builds from Unsloth and llama.cpp. The model page lists it under Apache-2.0 with 136 downloads and 39 likes.
The oversight argument
Looping is also sold as a safety property: reasoning inside a token is reasoning a monitor cannot read. In a variable-loop variant, the no-CoT time horizon rose from 16 seconds at one pass to 37 at three, chains of thought grew shorter but "stayed legible", and the model "didn't trend higher on tests of agentic sabotage as loops were added".
What looping cannot do is store facts. The vendor says its AA-Omniscience score is in line with models of its size, and its limitations section concedes 50.0% on MultiChallenge and 33.4% on ComplexConstraints, behind Qwen3.8-27B on both.
Inkling 2 Mini is an architectural bet rather than a new size class: same weights, more passes, about 20GB on disk. The numbers are the company's own, on charts it drew, footnoted to say the setups differ.


