Inkling 2 Mini on NDIF
The Looped Lens Challenge
Can you surface what Inkling 2 Mini computes but doesn't write?
Competition opens Nov 23
The challenge
A lens reads an activation from inside a model and says, in tokens, what the model is computing at that point. The logit lens and its relatives all read the same thing: the residual stream, the running sum that every layer of an ordinary transformer adds to.
Inkling 2 Mini is a small open-weights model from Thinking Machines. It loops its middle layers, so it does more computation on each token instead of writing more tokens. It also has no single residual stream: each layer reads its own mix of what earlier layers wrote. So what happens to the lenses we use to understand and monitor a model? Does a Jacobian lens still work the same way?
This challenge asks the community to find out: build the lens that best reads Inkling 2 Mini. It runs on NDIF, which hosts the model, and it is open to anyone with an NDIF key. You write an experiment on your own machine with nnsight, and NDIF runs it on the model.
Inside Inkling 2 Mini
Inkling 2 Mini has 24 layers of weights and applies 48 to every token. The middle twelve run three times, with the same weights on each pass.
- Prelude and coda, run once
- Core, pass 1
- Pass 2
- Pass 3
The nnsight blog post Opening the box of Inkling 2 Mini explains the loop and the blocks, with code to read them on NDIF.
Why it matters
Looped transformers are getting attention as an architecture that performs well, and as a subject for interpretability. Inkling 2 Mini is a major one: it runs on a laptop, and it is performant enough to power workflows in coding, agentic work, computer use and design.
At the same time, frontier models can do more and more of their reasoning inside a single forward pass, with no chain of thought to read: see the reports from Redwood Research, Epoch AI and the Alignment Forum. As part of this release, Thinking Machines studied the effects of looping on chain-of-thought monitorability for Inkling 2 Mini and is publishing the results, along with a second model built for studying looping.
Reasoning that a model does not write down can only be read from its activations. Lenses do that, and they have had little testing on looped models.
The competition
The goal is to crowdsource two things: new tools for working on this class of model, and general insights into Inkling 2 Mini itself. You enter by writing a lens. These are the rules.
- Submission
- A Python script, limited in size, with your lens artifacts sent as files or as links to files.
- Input
- Your program receives one prompt from our evaluation set.
- Read locations
- You choose N layers to read on each prompt, so picking good layers is part of the task. Readouts are taken at a single token position per prompt. You can choose that too, though it should probably be the final token.
- Output
- The top K ranked tokens at each of your N read locations: N × K tokens for the prompt, labelled with the layer they came from.
- Scoring
- The readouts for a prompt are collected into one list, and the list is given to a judge.
- Phases
- The competition will probably run in several phases of increasing complexity. Phase one is single-token readouts on relatively simple tasks, in the style of workspace-bench.
- Entry
- Anyone can enroll. You need an NDIF key and a registration.
- Dates
- Phase one opens on November 23. Its closing date will be announced here.
Take part
Competition opens Nov 23

