🇺🇸 English 🇰🇷 한국어 🇯🇵 日本語
Home › AI & Machine Learning
AI & Machine Learning

Inkling 2 Mini Loops Its Layers for Laptops

Thinking Machines Lab released Inkling 2 Mini on Nov. 20, a 26.4B-parameter open-weights looped transformer under Apache 2.0 that runs on machines with 32GB of memory. The lab also published a variable-loop research model to study how looping affects reasoning and chain-of-thought monitoring.

Thinking Machines Lab on Nov. 20 released Inkling 2 Mini, an open-weights model with 26.4 billion parameters that runs the middle half of its layers three times for every token. The lab says the roughly 20GB model is small enough to run offline on a laptop and is publishing the weights on Hugging Face under the Apache 2.0 license, along with a research variant whose loop count can be changed.

A laptop on a desk with a terminal window running code
File photo: a laptop running code in a terminal. Inkling 2 Mini is built to run locally on machines with 32GB of memory. · 사진: shabdaly / Flickr(CC BY 2.0)

Key takeaways

  • Inkling 2 Mini has 26.4B total parameters and 3.4B active, or 5.4B weight activations per token once looping is counted.
  • The lab reports 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, the top scores it lists among open models under 100B parameters.
  • A companion research model lets outsiders test how one to five loops affect reasoning and chain-of-thought monitoring.

How the Inkling 2 Mini looped transformer works

The model has 24 layers split into four blocks of six. According to the release post, the first and last blocks run once, while the two middle blocks share weights and run three times, for 48 layer applications per token from 24 stored layers.

The lab says looping the middle 12 layers beat looping the last 12 on 27 of 28 evaluations. In a scaling test, a non-looping version cost 7.5% less compute to train but needed 140% more inference compute to match the looped model's loss on 512,000-token agentic traces.

Running Inkling 2 Mini on a laptop

Most weights were trained and ship in 4-bit NVFP4, with the last four layers in 8-bit MXFP8, so there is no conversion step. The model card lists Apple silicon machines with at least 32GB of memory, a single NVIDIA DGX Spark, or GPUs with 32GB or more of VRAM. It supports a 1 million-token context window and accepts text and images.

Supported runtimes include Ollama, vLLM, SGLang, Unsloth, NVIDIA Dynamo and TokenSpeed, with quantized GGUF files for llama.cpp. The model is also available for inference and fine-tuning on the lab's Tinker platform.

Benchmarks and limits

All scores are the company's own. Inkling 2 Mini posts 48.2% on DeepSWE v1.1, against 42.2% for Qwen3.8-27B, and 79.2% on SWE-bench Verified versus 77.8%. The two tie at 20.1 on the Artificial Analysis Index v5.0.

The lab says looping improves reasoning but not stored knowledge. The model trails Qwen3.8-27B on multi-turn chat (50.0% on MultiChallenge) and on ComplexConstraints v1.1 (33.4%). The model card also warns it may delete or overwrite files without confirmation and recommends running agents in a sandbox.

Context: does looping hide reasoning?

Looping lets a model compute more without writing its steps out, which could weaken chain-of-thought monitoring. Using the variable-loop model, TML-d24-varloop-research, the lab found its no-chain-of-thought time horizon rose from 16 seconds at one pass to 37 seconds at three, then roughly leveled off.

On GPQA Diamond, average reasoning fell from 3,703 tokens at one pass to 1,661 at five, while sentence length and dictionary-word share stayed the same. In LinuxArena sabotage tests, the lab saw no clear trend as loops increased. It calls the findings preliminary.

What's next

Inkling 2 Mini follows Inkling 2 Small, a 594B model with 17B active parameters released Nov. 5. The lab says it is working with the National Deep Inference Fabric to give interpretability researchers easier access to the looped models.

FAQ

Is Inkling 2 Mini open source?
Its weights are on Hugging Face under the Apache 2.0 license, which allows commercial use.

What hardware does Inkling 2 Mini need?
About 32GB of memory: an Apple silicon Mac with 32GB or more, an NVIDIA DGX Spark, or a GPU with 32GB or more of VRAM.

What is a looped transformer?
A model that reuses the same layers several times per token, adding computation depth without adding stored parameters.

References

#Thinking Machines Lab #Inkling 2 Mini #looped transformer #open weights #local LLM #Hugging Face #Apache 2.0 #chain of thought #monitorability #NVFP4
S
Seung Jung
AI Newsway · Developer

Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.

More by this reporter ›
Copyright © AI Newsway All rights reserved. Unauthorized reproduction, redistribution, and AI training prohibited.
Share
f X K

Comments 0

Comments are currently disabled.

Be the first to comment.