Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX Spark
A.I. / news
Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX Spark
The model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.

Thinking Machines Lab released Inkling 2 Mini on Hugging Face on November 20, 2026, a 26.4B-parameter mixture-of-experts model under Apache 2.0 that its card says runs on a single NVIDIA DGX Spark.
The weights are downloadable and the card does not mention a gating form. The lab also offers the model through its Tinker API, according to the card.

What the Inkling 2 Mini card specifies
The card describes a 24-layer looped transformer. The middle 12 layers run three times per token, and each token is routed to 8 of 128 experts.
Total parameters are 26.4B, with 3.4B active. Counting the looped passes, the card puts effective active parameters at about 5.4B per token.
The context window is up to 1M tokens. Inputs can be text or images, and output is text.
| Spec | Inkling 2 Mini card |
|---|---|
| Total parameters | 26.4B |
| Active parameters | 3.4B (about 5.4B with loops) |
| Context window | 1M tokens |
| Checkpoint | about 20GB, mixed NVFP4 and MXFP8 |
| Licence | Apache 2.0 plus acceptable use policy |
Hardware and the GGUF builds
The card says the model runs on one DGX Spark, on Apple silicon with 32GB or more of memory, or on other GPUs with roughly 32GB or more of VRAM.
Unsloth, a separate company, published GGUF quantisations of the model. Its page lists 238,152 downloads in the last month. Its own tests against its UD-Q8_K_XL reference (22.87GB) show how much each smaller file costs in fidelity.
- UD-IQ1_S7.76 GB
- UD-Q3_K_XL13.63 GB
- UD-Q4_K_M16.61 GB
- UD-Q8_K_XL22.87 GB
Source: Unsloth model page on Hugging Face, accessed 2026-11-21
The Q4_K_M build keeps the same top token as the Q8 reference 97.9% of the time, Unsloth reports. The smallest build, UD-IQ1_S, matches 89.4% of the time.
Unsloth also flags two behaviours. The model can get stuck in long reasoning loops, especially when debugging code, and setting reasoning effort to minimum does not fully switch reasoning off. The card recommends capping output tokens and agent turns.
Benchmarks without an evaluator
The card compares Inkling 2 Mini with Gemma 4 31B, Qwen3.8-27B and Nemotron 3.5 Super VL. Some rows cite Artificial Analysis and Vals indices, and agent harnesses such as mini-swe-agent and OpenHands.
The card does not state who ran most of those evaluations. Treat every figure on it as vendor-supplied until a third party reproduces it.
Training data is described as public sources, third parties, internal generation and synthetic data, spanning text, images and video. The card links to separate documentation. For how another lab handles synthetic outputs, see OpenAI's distillation safeguards.
Where it sits in the Inkling line
The lab's first open-weights model, Inkling, shipped in July with 975B total parameters and 41B active under Apache 2.0, Simon Willison wrote. That post also named Inkling-Small, a 276B model with 12B active, whose weights were to follow once testing finished.
Inkling 2 Mini is roughly one thirty-seventh of the original's total size. Another small model on the lead sheet is Dettato's 20M-parameter audio-to-MIDI chain.
The Hugging Face page showed 2,127 likes and 7,944 downloads on the main repository as of November 21. The lab has not published a date for a larger Inkling 2 release.
Sources
- 03Inkling
More in A.I.
- 01Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.
- 02Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings FigureAnthropic and OpenAI each reported token reductions for the same defer_loading flag. Google's November 19 post reports none.
- 03OpenAI Says New Safeguards Cut Reasoning Extraction From 54.0% to 0.2%The figures are from OpenAI's own red-team evaluation, and the post names DeepSeek and Moonshot AI as distillers of U.S. models.
- 04Frostbite, an AI-Designed CPU, Claims 4.59x Efficiency Over an Unnamed BaselineArchitect Labs reports pre-silicon results from a 7 nm layout and has not said the chip has been fabricated.