Featured:MeloLab– Create songs and background music with AI.

Inkling 2 Mini

Inkling 2 Mini is a 26.4B-parameter looped Mixture-of-Experts model with open weights, ~20GB memory footprint, and strong coding and agentic benchmark scores.

Type:
AI Tool
Inkling 2 Mini product preview
Contents

What is Inkling 2 Mini?

Inkling 2 Mini is a 26.4B-parameter looped Mixture-of-Experts model with open weights, ~20GB memory footprint, and strong coding and agentic benchmark scores.

Updated ·

Inkling 2 Mini: Looped MoE Open-Weights Model

Announcement date:

Inkling 2 Mini is an open-weights Mixture-of-Experts language model from Thinking Machines Lab that reuses its middle layers: the model runs those layers three times for every token. It has 26.4B total parameters but only 3.4B active, and its total memory footprint is about 20GB, so it is designed to run on local devices such as consumer laptops and desktop AI computers. It is available on Tinker for inference and customization, and it leads open-weights models under 100B total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%).

Social announcement artwork introducing Inkling 2 Mini
Publisher announcement artwork; not a product interface screenshot.

What Inkling 2 Mini is

Inkling 2 Mini is a looped transformer: rather than passing each token through every layer once, it runs the middle layers of the model three times per token. The architecture has 24 layers split into four blocks of six; the two middle blocks run three times, while the first and last blocks each run once.

It is also a Mixture-of-Experts model. It carries 26.4B total parameters, activates 3.4B, and accounts for 5.4B weight activations per token because of the looping. That combination is what keeps the model small enough to deploy locally while still giving it more effective depth.

  • Looped transformer: middle layers run three times per token
  • 24 layers in four blocks of six; first and last blocks run once
  • 26.4B total parameters, 3.4B active, 5.4B weight activations per token
  • Mixture-of-Experts design

How the architecture and training pipeline fit together

The model was shaped partly by pruning. At one point during midtraining, two-thirds of the experts were removed, leaving the 128 most-used experts in each layer to cut the total parameter count. After pruning, the model trained on 1T tokens distilled from a larger internal Inkling model to recover pre-pruning performance.

Post-training reinforcement learning totaled 2.2T tokens over 2,000 steps. Thinking Machines also trained a separate variable-loop research model, TML-d24-varloop-research, with between one and five looping passes and released it publicly on Hugging Face, which is useful if you want to study how loop count affects behavior rather than deploy the production model.

  • Two-thirds of experts pruned during midtraining, keeping the 128 most-used per layer
  • 1T distillation tokens from a larger internal Inkling model after pruning
  • RL post-training: 2.2T tokens over 2,000 steps
  • Separate variable-loop research model (1–5 passes) on Hugging Face

Memory footprint and native quantization

The total memory footprint is about 20GB, which is the headline practical number: it is small enough for consumer laptops and desktop AI computers rather than requiring a multi-GPU server.

Most of the weights were trained and are deployed natively in NVFP4, with the last four layers in MXFP8. Because the low-precision formats are native rather than bolted on afterward, the ~20GB figure reflects how the model is actually shipped, not an aggressive post-hoc quantization.

  • ~20GB total memory footprint
  • Runs on consumer laptops and desktop AI computers
  • Most weights natively NVFP4
  • Last four layers in MXFP8

Where it leads and where it trails

Among open-weights models with fewer than 100B total parameters, Inkling 2 Mini leads on DeepSWE v1.1 at 48.2% and Vals Index v2.2 at 46.5%. Those results point to coding and agentic or index-style evaluation as its stronger areas.

It is not the strongest small model for multi-turn chat or instructions with many constraints. It scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, behind Qwen3.8-27B on both. If your workload is long conversational threads or dense constraint lists, that gap is worth weighing before you standardize on it.

  • Leads open-weights models under 100B total parameters on DeepSWE v1.1 (48.2%)
  • Leads the same group on Vals Index v2.2 (46.5%)
  • MultiChallenge: 50.0%, behind Qwen3.8-27B
  • ComplexConstraints v1.1: 33.4%, behind Qwen3.8-27B

Access and getting started

Inkling 2 Mini is released with open weights and is available on Tinker today for inference and customization. A practical starting sequence is to try it on Tinker first to gauge output quality on your own prompts, then move to local deployment if the ~20GB footprint fits your hardware.

If you plan to run it locally, check that your runtime supports NVFP4 and MXFP8 weights before downloading, since most of the model ships in those formats. For research on loop behavior, the variable-loop model on Hugging Face is the more appropriate starting point than the production checkpoint.

  • Open weights release
  • Available on Tinker for inference and customization
  • Confirm NVFP4 and MXFP8 support in your local runtime
  • Use the variable-loop research model for loop-count experiments

Pricing status

The source material describes Inkling 2 Mini as open weights and available on Tinker, but it does not state pricing for Tinker inference or customization, and no license terms or commercial-use conditions are given. Treat cost and licensing as unknown until you confirm them directly with the provider.

  • Open weights confirmed
  • Tinker availability confirmed
  • Tinker pricing not stated in the source
  • License and commercial-use terms not stated

Questions

How does the looping in Inkling 2 Mini actually work?

The model has 24 layers in four blocks of six. The two middle blocks run three times for every token, while the first and last blocks run once. That reuse of middle layers is what the term looped transformer refers to, and it is why the model reports 5.4B weight activations per token despite 3.4B active parameters.

Can Inkling 2 Mini run on a laptop?

The total memory footprint is about 20GB, and the source describes it as able to run on local devices such as consumer laptops and desktop AI computers. In practice you also need a runtime that supports its native NVFP4 weights and MXFP8 final layers.

Is Inkling 2 Mini better than Qwen3.8-27B?

It depends on the task. Inkling 2 Mini leads open-weights models under 100B total parameters on DeepSWE v1.1 and Vals Index v2.2, but it trails Qwen3.8-27B on MultiChallenge (50.0%) and ComplexConstraints v1.1 (33.4%). Choose based on whether your workload is closer to coding and agentic evaluation or to multi-turn chat and constraint-heavy instructions.

Original sources

  1. thinkingmachines.ai — Original source 1 ↗

Tool Info

Has this tool changed? Help us keep it up to date

See also: related tools and guides

Loading reviews…

Inkling 2 Mini Reviews

Not rated yet0 user reviews

Do you recommend Inkling 2 Mini? Leave a comment below!

Your rating
0 / 2000

Loading reviews…

Categories

FOR PRODUCT BUILDERS

Built something useful? Help people find it.

Start with your website. Add your logo, screenshots and product details to a listing you can keep up to date.

Verify your website badge for a free listing.