Laguna-XS.2-jlens

Laguna-XS.2-jlens

Holds Jacobian lens files that compare a bf16 and an NVFP4 copy of Laguna-XS-2.1, plus a version comparison with Laguna-XS.2. Research files for people who study model internals.

Made by PrimeIntellect on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds research files, not a chat model. They are Jacobian lenses fitted on poolside's Laguna-XS-2.1, a mixture-of-experts model with about 33B parameters, 40 layers and 256 experts. One lens comes from the bf16 checkpoint and one from its NVFP4 twin, so the differences point to expert quantization.

Who it suits: Researchers who study how quantization or a new model version changes a model's inner structure. It suits people who read reports and load PyTorch files.

Industry | Who uses it | How they use it | Example
AI research | Interpretability researchers | Compare lens fits across two precisions | Check how far NVFP4 experts shift the layers
Model engineering | Quantization teams | Judge what a 4-bit expert format changes inside a model | Read the A/B report before shipping a quantized copy
Model review | Evaluators | Compare two releases of the same model | Look at the Laguna XS.2 and XS-2.1 comparison

Anyone who wants to chat with a model or generate text can skip it, because it contains analysis files only.

Where to use it:

• PyTorch: the lens files are .pt checkpoints.
• Python scripts or notebooks, for loading and comparing lenses.
• The author's open-jlens-data repository: it holds the code that fitted the lenses.
• Windows, macOS or Linux: the download steps run on all three with Python.

Use case | Who | What they get back
Quantization A/B | Researchers | Two merged lenses fitted on the same 200 prompts
Stability check | Researchers | Eight per-GPU shard lenses per arm, plus convergence files
Version comparison | Researchers | A second fit for Laguna-XS.2 and a comparison with XS-2.1

How to set it up: What you need first:

• Python with pip.
• PyTorch, to open the .pt files.
• Optional: the open-jlens-data repository, for the fitting and analysis code.

1. Install the Hugging Face hub package with pip.
2. Download this repository with the hf command.
3. Open the report in the results folder first.
4. Load the lens files in PyTorch when you want to compare them.

pip install -U huggingface_hub
hf download PrimeIntellect/Laguna-XS.2-jlens

How to use it: Start with the report and the summary file in the results folder. Then compare the two merged lenses, one per precision. The shard and convergence folders let you check how stable each fit is. The xs2 folder holds the same fit for the older Laguna-XS.2.

Title: Files and folders the card names
Tool | What it does
laguna_bf16_jacobian_lens.pt | Merged lens from the bf16 checkpoint
laguna_nvfp4_jacobian_lens.pt | Merged lens from the NVFP4 twin
shards folders | Per-GPU lenses, 25 prompts each, with the same prompts in both arms
convergence folders | CSV files that track how each shard's fit converges
results folder | The summary file, analysis arrays, figures and the written report

Example: You open the written report, then load both merged lenses and compare them layer by layer. The card says the NVFP4 change measures 0.008 on a 1-CKA scale, close to the 0.007 you get from prompt noise.

For newcomers and experts: A lens is a tool for reading what a model's layers are doing. Here the same model is read twice, once at full precision and once with compressed experts, to see whether compression changed its inner structure. A good first step is to read the report.

The two lenses use identical prompts, 200 from wikitext-103 at length 128, across all 39 source layers. The non-expert tensors match bit for bit, so the gap comes from expert quantization. Moving from XS.2 to XS-2.1 changes the lens about 43 times more than NVFP4 does, yet the depth structure survives in both.

Take 200 identical prompts
Fit one lens on bf16 and one on NVFP4
Compare layers with CKA
Read the gap against prompt noise

Good to know: The card ranks three sources of change on the same prompts, measured as 1 minus CKA.

Title: Change on identical prompts, 1 minus CKA
Prompt noise: 0.007 units
NVFP4 quantization: 0.008 units
Model version: 0.362 units

• Scope: the files cover Laguna XS-2.1 and Laguna XS.2 from poolside.
• Caveat: the card says the finding may not carry over to larger models, because Laguna's experts are smaller than those in the Inkling lens.
• License: apache-2.0
• Cost to run: no API key, account or paid service named.
• Data it reaches: files you download from Hugging Face. The fit used text from wikitext-103.
• Not for: chatting, text generation, or running as a model.

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 0
• Likes: 7
• License: apache-2.0
• Quantized versions: 1
• Tags: interpretability, jacobian-lens, nvfp4

Links

• Model: https://huggingface.co/PrimeIntellect/Laguna-XS.2-jlens
• Files: https://huggingface.co/PrimeIntellect/Laguna-XS.2-jlens/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download PrimeIntellect/Laguna-XS.2-jlens

Then follow the usage in its model card. Nothing in this kit runs by itself.