# Is Inkling 2 Mini SOTA?

Pending benchmarks. Announced Nov 20, 2026; no independent results yet.

As of Nov 22, 2026, Inkling 2 Mini is pending benchmarks on isitsota: announced Nov 20, 2026, with no independent results on the boards we track yet.

Thinking Machines' smaller Inkling 2 sibling: a multimodal looped-transformer MoE with 26.4B total parameters and 3.4B active (5.4B effective active per token with looping), a 1M-token context window, and a stated memory footprint of about 20GB. Official model card dates the release to November 20, 2026. Ungated Apache 2.0 weights are on Hugging Face; the model is also on Tinker and the Tinker Playground.

Canonical: https://isitsota.com/m/inkling-2-mini
Last updated: 2026-11-20
Tracking since: 2026-11-20

## Announced facts

| Fact | Value |
| --- | --- |
| Lab | thinky |
| Announced | Nov 20, 2026 |
| API | Live |
| Open weights | Published · Ungated open weights on Hugging Face at thinkingmachines/Inkling-2-Mini. Official card: a single checkpoint of about 20GB in mixed NVFP4 and MXFP8. Official blog: quantized GGUF versions are available via Unsloth and llama.cpp. Official blog also releases a research-specific variable-loop variant alongside this model (Thinking Machines blog and model card) |
| Parameters | 26.4B total / 3.4B active |
| Architecture | Official model card: a 24-layer decoder-only multimodal looped transformer with sparse MoE feed-forward layers; the middle 12 layers run three times with shared weights (48 layer applications per token). Each token is routed to 8 of 128 experts. Official blog: 26.4B total parameters, 3.4B active, and 5.4B weight activations per token to account for looping. Most weights are trained and deployed in NVFP4, with the last four layers in MXFP8 (Thinking Machines blog and model card) |
| Context | Official model card: a context window of up to 1M tokens (Thinking Machines model card) |
| Modalities | Official model card: text and image inputs, text output (Thinking Machines model card) |
| Hosted access | Official blog and model card: available on Tinker for inference, fine-tuning and customization, and for chat on the Tinker Playground. Official card also lists local inference on Hugging Face, SGLang, vLLM, TokenSpeed, NVIDIA Dynamo, Unsloth (including quantized GGUF), llama.cpp (quantized GGUF), and Ollama and Exo Labs on Apple silicon, and names 1x NVIDIA DGX Spark and Apple silicon devices with 32GB or more memory among supported hardware (Thinking Machines blog and model card) |
| License | Apache 2.0 (Thinking Machines model card) |
| Related | Smaller sibling of Inkling 2 Small; separate from Inkling and Inkling Small |

Tags marked unsure are not stated on the lab's own pages.

## What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

- [x] Publicly announced with a source
- [x] Usable by the public (API or weights)
- [ ] Listed on an independent board we track
- [ ] Enough results for a rank in a category
- [ ] Holds across two snapshots (crowns only)

## What the lab claims

Self-reported. Shown for context; not used in any verdict.

- Lab claim: self-reported 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2 (official Thinking Machines blog and model card).

## Sources

- LAB: [Introducing Inkling 2 Mini (Thinking Machines)](https://thinkingmachines.ai/news/inkling-2-mini/) (2026-11-20)
- LAB: [Inkling 2 Mini (Thinking Machines model card)](https://huggingface.co/thinkingmachines/Inkling-2-Mini) (2026-11-20)

Tracking continues until an independent board lists it.

## Related

- Lab: [thinky](https://isitsota.com/labs/thinky.md)
- Compare when it lands (unscored): [Inkling 2 Small](https://isitsota.com/m/inkling-2-small.md)
- All pending releases: https://isitsota.com/releases.md
- JSON: https://api.isitsota.com/v1/models/inkling-2-mini
