What is Inkling 2 Mini?

Inkling 2 Mini is Thinking Machines Lab's open-weight looped transformer, released on 20 November 2026. It has 26.4B total parameters, 3.4B active, and a context window of up to 1 million tokens. Official docs list a memory footprint of about 20GB, Apache 2.0 weights on Hugging Face, local runtimes including Ollama and Unsloth, and Tinker prices for training and sampling.

What Inkling 2 Mini is for

Thinking Machines Lab presents Inkling 2 Mini as its smallest Inkling 2 model: a looped transformer that runs on a laptop and, the company says, excels at agentic coding. The official family page and announcement describe a model that runs its middle layers three times per token, which the lab says lets it apply more reasoning with fewer written tokens. Official specs are 26.4B total parameters, 3.4B active, and 5.4B weight activations per token to account for looping.

The official Hugging Face card, which is also the place to check common questions about size, Hugging Face weights, Ollama, Unsloth, GGUF, and benchmarks, says the model accepts text and image inputs and generates text, with a context window of up to 1 million tokens. The official memory footprint is about 20GB, in mixed NVFP4 and MXFP8. The announcement reports a laptop demo on a MacBook Pro (M5 Pro, 48GB) in airplane mode.

Vendor-reported scores include 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, which the company says lead open-weight models with less than 100B total parameters. Treat those as official published figures, not independent tests.

A first path to try it

If you want to run it locally, the official model card lists Hugging Face, SGLang, vLLM, TokenSpeed, NVIDIA Dynamo, Unsloth including quantized GGUF versions, llama.cpp for GGUF, Llama App, and Ollama and Exo Labs on Apple silicon. The card recommends Apple silicon with 32GB or more memory, or other GPUs with 32GB or more of VRAM, and it warns that support for looped models in local runtimes is new and that speed varies.

If you want a hosted path, the official announcement says Inkling 2 Mini is available on Tinker for inference and customization, and the card adds the Tinker Playground for chat. A research-only variable-loop variant is also on Hugging Face as TML-d24-varloop-research.

This is a suggested official path, not a recorded test. After a first task, set a limit on output tokens and agent turns, as the official limitations section recommends, and sandbox any run that can edit or delete files.

Pricing, access, and limits

The official Hugging Face model card, checked on 21 November 2026, states the license as Apache 2.0, matching the README metadata (license apache-2.0) and the card's apache-2.0 tag. Weights are free to download there under that license. Confirm the live card if you need the full Apache 2.0 terms or the separate Acceptable Use Policy linked from the card.

The official Tinker pricing page, labeled as of 19 October 2026 and checked on 21 November 2026, lists Inkling 2 Mini training and sampling in USD per million tokens: 64K prefill $0.15, cached prefill $0.015, sample $0.50, train $0.75; 256K keeps the same prefill, cached, and sample rates, with train $1.50. Storage is $0.10 per GB per month. The serverless inference table on the same page is labeled as available for Inkling 2 Small, Inkling, and Inkling-Small, and it does not include a Mini row, so no serverless Mini rate is quoted here.

Official limitations include weaker factual recall than Inkling 2 Small, occasional long reasoning loops, lower reliability with tools than larger models, possible destructive actions without confirmation, and weaker quality outside English. Setting reasoning effort to its minimum does not fully disable reasoning, according to the card, because the middle layers still run three times.

What stands out

  • Official looped mixture-of-experts design: 26.4B total parameters, 3.4B active, and 5.4B weight activations per token because the middle layers run three times
  • Official context window of up to 1 million tokens, with text and image inputs and text output
  • Official open weights under Apache 2.0, plus local support for Hugging Face, SGLang, vLLM, Unsloth including quantized GGUF, llama.cpp GGUF, Ollama and Exo Labs on Apple silicon, and others listed on the model card
  • Official Tinker training and sampling prices, plus Tinker Playground chat, checked on the live Tinker pricing page on 21 November 2026
  • Thinking Machines reports that the model leads open-weight models with less than 100B total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%)

Before you choose

  • Official limitations include weaker factual recall than Inkling 2 Small, a risk of long reasoning loops, less reliable tool use than larger models, and possible destructive file actions without confirmation
  • The official Tinker serverless inference table checked on 21 November 2026 does not list Inkling 2 Mini, so this profile does not invent a serverless rate
  • Benchmark figures in this profile are Thinking Machines Lab's reported results, not independent measurements
  • This profile is not a hands-on test of Tinker or local runtimes such as Ollama or Unsloth

Plans & pricing

Current pricing on thinkingmachines.ai (opens in a new tab)

Sources & further reading

Based on the official documentation below. Features and access can change; check the provider before making a decision.

  1. Thinking Machines Lab: Introducing Inkling 2 Mini, dated 20 November 2026, rendered 21 November 2026 (opens in a new tab)
  2. Official Hugging Face model card for thinkingmachines/Inkling-2-Mini, rendered 21 November 2026 (opens in a new tab)
  3. Official Tinker pricing page, labeled as of 19 October 2026, rendered 21 November 2026 (opens in a new tab)
  4. Official Inkling family page, rendered 21 November 2026 (opens in a new tab)
Something out of date or incorrect?Report a correction

Was this page useful?

Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.

Share this page

Related reading