ModelMRI

Chrome DevTools for AI models and agents.

▶ Live demo · Open a .mri · Docs · Build log

Load any local model — LLM, VLM, or robot policy — and see inside it while it runs: what it attended to, which concepts fired, what happens when you turn one off, and exactly where your agent went wrong.

ModelMRI is an open-source, local-first interpretability and debugging tool for transformer language models, vision-language models, robot policies and LLM agents. It visualizes per-layer, per-head attention weights from a live forward pass, ranks attention heads by causal ablation scored with KL divergence, decomposes the residual stream with sparse autoencoders, steers generation along a feature direction, maps activations in any custom nn.Module, and records agent runs as an inspectable timeline. It runs on your own machine — no cloud, no account, no telemetry — and writes findings to a .mri file a colleague can open in a browser with nothing installed.

Python 3.10+, Windows / macOS / Linux, MIT licensed.

Hover any token — arcs show what it attended to. Every layer, every head.

pip install modelmri
modelmri serve          # open http://localhost:5900

It finds the models you already have — HF cache, plain folders, GGUF — before asking you to type anything.

What you can actually do with it

1. See what a token attended to

Type a prompt, watch it stream, then hover any token — arcs show which earlier tokens it looked at, scaled by attention weight, for any layer and head.

On GPT-2, the generated token " Paris" attends back to " capital" and " France". The information was always there. Nobody was looking.

2. Ask which heads actually mattered

144 heat maps and no reason to open any of them is a browsing tool. Rank heads zeroes each head in a layer, runs the model again, and measures how far the answer moves — so the dropdown arrives ordered and the top head is already selected.

gpt2 · "The capital of France is" · zero-ablation · bf16

Rank heads → L0 H7  KL 0.898   p(" the") 0.098 → 0.057
             L0 H10 KL 0.535
             L0 H9  KL 0.412

The setup line is not decoration. The same three heads on the same model score 0.784 / 0.543 / 0.415 in fp32 and 0.825 / 0.559 / 0.469 over a 261-token generation — a KL depends on the prompt, the dtype and the sequence, so a figure quoted without them cannot be checked by anyone.

A ranking costs n_heads + 2 forward passes; the whole model costs n_layers × n_heads + 2. That is the part that is portable — gpt2 is 146 passes, Qwen3-0.6B is 450. What a pass costs on your machine is not: measured on one RTX 4060 across sessions it moved between 12 and 71 ms for the same model, so the panel measures a layer on your machine and extrapolates from that rather than quoting a number from mine. One layer by default; the whole model only when told, with the estimate shown first.

Then ask what changes? on any ranked head and the panel subtracts the two runs — arcs in one colour where the model attends more without that head, another where it attends less. It opens at layer L+1, because removing a head cannot change its own layer's attention (that layer's input is unchanged), and a zero result says so rather than showing you an empty canvas.

Both sides are forward passes over the same token sequence, never two generations — sampling diverges, and chat templates insert 0, 8 or 29 leading tokens depending on the model, so subtracting two generations would align token 5 of one against token 5 of a different sentence.

It reports what it measured and nothing more. These are not each head's share of the prediction — on gpt2 layer 0 the twelve per-head scores sum to 1.995 while ablating the whole layer gives 0.208 — and the ranking depends on what a removed head is replaced with, so the baseline is named on screen and both are offered. head_dim is read from the model rather than computed as hidden_size // n_heads, which is wrong by 2× on Qwen3 and would rank half-heads confidently.

3. Ask where in the model the answer is decided

Ablation says what mattered. It cannot say where the thing is. Patching takes two prompts that differ in one fact, moves an activation from the run that knows the answer into the run that does not, at every (layer, position), and reports how much of the difference comes back.

Blue recovered the clean answer, red pushed it further away. Ringed cells were tested against chance.

clean    "The Eiffel Tower is located in the city of"   ->  " Paris"
corrupt  "The Colosseum is located in the city of"      ->  " Rome"

Three grids, because where and through what are different questions — and they disagree. Measured on the same pair across three architectures:

The MLP peak sits on a subject token in an early layer in all three — um, os and osseum are all pieces of "Colosseum" — while the attention peak sits on the last token, late. Early MLP writes the fact; late attention carries it to where the prediction is made. The residual grid contains both and shows only the destination.

The score is signed, and it is the one ranking here that is not a KL: a patch can push the answer further away, and 5 of 132 sites did. It is also not capped at 1.0 — a single site can overshoot, and gemma-3-270m-it reads 1.010.

Each of the strongest sites is run again against eight same-norm random draws at the same site, not one, because one is a coin flip: at a single site the draws ran from −2.038 to +0.616 against a real recovery of +0.427.

Most casually-written pairs are refused, and both failures are invisible unless you are told — the prompts must tokenize to the same length (2 of 8 natural minimal pairs did not) and must predict different tokens (2 of 3 did not, making the denominator exactly 0). Both refusals print what to change.

4. Find a concept and turn it off

Load a sparse autoencoder and ModelMRI shows the human-interpretable features firing on every token. Click one, drag the slider, and run a deterministic A/B:

gpt2 · layer 8 · jbloom/GPT2-Small-SAEs-Reformatted · FVU 0.0010 · 60.5 features/token

prompt                The Eiffel Tower is located in the city of
baseline               Paris, France.
feature #5856 @ -40    London's central London borough.

#5856 is the top-firing feature on the final prompt token — activation 35.55, the one the panel already has selected — so this example is reachable by following the instructions above rather than by knowing which number to type. Same prompt, greedy decoding, no prompt tricks. We reached into layer 8 and turned the concept down. Clearing the steer restores the baseline byte for byte.

The SAE checks itself before it shows you anything. An SAE fed the wrong activation convention does not error — it returns features, in the right shape, with plausible magnitudes, for a vector it never saw. So ModelMRI measures which convention actually reconstructs (centered along d_model or not, b_dec subtracted from the input or not), reports the fraction of variance unexplained, and refuses to plot anything when no convention reconstructs. On the default SAE that is the difference between 60.5 features firing per token and 7,491, and between an FVU of 0.0010 and 13,579.

5. Find the step where your agent died

Two lines of modelmri.record around any agent run gives you a timeline: LLM calls, tool calls, subagents, each as a block. The failure glows. Click it for the exact input, output, tokens, and error.

from modelmri.record import trace, step

with trace("fix-failing-tests"):
    step("llm_call", name="plan", input=prompt, output=answer, tokens_in=912)
    with step("subagent", name="auth-fixer"):
        step("tool_call", name="pytest", output="17 passed")

Or instrument automatically: modelmri.record.instrument_anthropic().

6. Look inside a robot policy

This is the part nobody else ships. ModelMRI loads the vision tower of the real SmolVLA checkpoint and runs actual robot-camera frames through it, painting each image patch's attention back onto the frame. Scrub an episode, run the policy, drag the layer slider.

Measured on PushT frames — share of attention mass in the top 5% of patches:

Early layers look everywhere; deep layers lock on. No robot hardware required — it reads public LeRobot datasets straight from disk.

7. Debug a model you trained yourself

Everything above is transformer-shaped. This isn't. Point ModelMRI at your own nn.Module — an MLP, a small CNN, whatever you're training — and get a layer-by-layer map of one real forward pass.

# my_net_adapter.py — the whole contract
def load():
    model = MyNet()
    model.load_state_dict(torch.load("checkpoints/best.pt", map_location="cpu"))
    return model

Dead units, saturated activations, and the first layer where a nan appears — statistics exclude non-finite values on purpose, so one bad number can't turn every row below it into nan and hide where it started.

A state_dict alone is refused, with the reason: it's weights without an architecture, and guessing one would produce a map that looks authoritative and describes a network you never trained.

8. Send someone the finding, not the model

You found the head. Now show a colleague — who does not have your GPU, your prompt, or 8 GB of spare disk.

That writes one 54 KB file holding the tokens, the attention, the generation and your note. No weights — it's an observation, not a checkpoint.

The recipient opens it at the viewer — nothing installed, nothing uploaded, the file is read in their browser.

Locally it's the same page, served from the package by the standard library:

modelmri open gpt2.mri     # ~0.3s — no torch, no model, no GPU

If you were sent several and want to know which is which, inspect prints one without opening anything:

modelmri inspect gpt2.mri

gpt2.mri — 41.8 KB
  model         gpt2
  size          124M parameters
  ran on        cuda · float32
  recorded      2026-08-11T17:55:27+00:00 by ModelMRI 0.9.0
  note          the head that carries the city

  tokens        11 (10 prompt)
  shape         12 layers x 12 heads
  attention     144 maps
                every layer and head
  patching      attn, mlp, resid
    clean       The Eiffel Tower is located in the city of
    corrupt     The Colosseum is located in the city of

  prompt        The Eiffel Tower is located in the city of
  answer         Paris

--json gives the same summary as a document, with the prompt and generation untruncated. Both take the same ~0.2s as open: a .mri is gzipped JSON, so nothing here needs torch.

A recording carries the activation-patching trace too, when one was run — so the causal finding travels with the file, not just the attention. Open a .mri that has one and the patching panel draws it, marked as recorded rather than measured on your machine. A recording that carries none says so instead of offering a button that can only refuse: patching means running the model again with an activation replaced, and a .mri holds activations rather than weights.

Every panel reads a recording through the same calls it uses for a live model, so the arcs, the layer/head dials and the token strip all behave normally. The status pill says replay and the footer says recorded, not live, so it can never be mistaken for your own run.

The browser viewer and the Python tool are checked cell-for-cell against the same file on every change (tests/viewer_check.py) — a viewer that renders a slightly different matrix would be worse than no viewer, because nothing on screen would say so.

Every command

Nine of them, and only one needs a GPU. Everything below reads what is already on your disk — no model is downloaded and no server is started unless the command says so.

The two you will use most

modelmri models        # what have I got, and why won't that one open?
modelmri serve         # look inside one of them

Sending someone a finding

modelmri inspect gpt2.mri     # what is in this file?
modelmri open gpt2.mri        # show me

A .mri is one analysis without the model — attention, the logit lens, the generation, and the activation-patching trace if you ran one. It is how you show a colleague the head you found without asking them to download 8 GB.

When something is wrong

modelmri doctor        # can this machine run it at all?
modelmri where         # where did it put my stuff?

Moving where things go

Install

pip install modelmri              # core: playground, attention, features, steering, agents
pip install "modelmri[vla-lite]"  # + robot datasets (av, pyarrow, pillow)
pip install modelmri-record       # just the agent recorder — stdlib only, an 10.9 KiB wheel
modelmri doctor                   # what this machine can and cannot run, measured
modelmri serve

Will it run here? modelmri doctor measures your machine and says so — OS, cores, RAM, free disk, the torch build, the accelerator it found and its precision, and roughly how large a model fits. It exits non-zero when something would stop a load, so it is scriptable. modelmri serve prints the one-line version at startup.

  accelerator NVIDIA GeForce RTX 4060 Laptop GPU (cuda)   vram 8.6 GB   bfloat16
  Models up to roughly 3.2B parameters should fit.

Every figure is read off the machine at the moment you ask, and a number that cannot be determined says "could not measure" rather than being invented. Note this is a first-run check rather than an install-time one, deliberately: a wheel is an archive and pip does not execute code from it, so there is nowhere honest to put a check during pip install. It is also the better place for it, because the same machine can be perfectly able to open a shared .mri and unable to load a 7B model — and those are different questions.

From source:

git clone https://github.com/muhammadmahadazher/ModelMRI && cd ModelMRI
cd frontend && npm ci && npm run build && cd ..
uv sync && uv run modelmri serve

Models. Search HuggingFace, pick from what's already cached on your machine, or switch to Ollama and pull any model by name. (Ollama gives you text only — internals need a HuggingFace model, and ModelMRI says so rather than pretending.)

Nothing downloads by surprise. Every row shows its size before you click, and a download that cannot fit your disk is refused with both numbers rather than started. One that dwarfs your GPU asks first. Whatever is running, Stop actually stops it — the fetch happens in a child process precisely so it can be killed, and the half-written blobs are cleaned up. This exists because a click once began fetching 1.5 TB onto an 8 GB laptop with no way out but killing the server.

GPU when you have one. NVIDIA, AMD, Intel and Apple silicon are detected automatically and the badge explains its choice — including the common case where torch was installed as a CPU-only build, where it prints the exact command to fix it. CPU works fine too; a 0.5B model streams in a couple of seconds.

API

The UI is a client of a plain HTTP API — script against it directly.

Status

Honest limits

• Attention needs eager attention. SDPA and FlashAttention never materialize the weights, so ModelMRI loads models with attn_implementation="eager". Slower, but it's the only way to see anything.
• SAE features need an SAE that exists. They are trained per model, and public ones exist for only a handful — this build knows of four repositories — so there is none for most of what you will load, and no amount of code makes one appear. ModelMRI offers the one that matches your model, says plainly when there is none, and falls back to a logit lens, which needs nothing but the model.
• Custom models get a layer map, not attention. Attention and SAE features need a transformer; for an arbitrary nn.Module ModelMRI shows shapes, activation statistics and pathologies. Loading an adapter runs your Python — see SECURITY.md.
• VLA mode is the perception half. SmolVLA's vision tower is real and loaded from the real checkpoint; the action expert needs lerobot, whose torch/numpy pins conflict with the core runtime, so it lives behind an opt-in extra rather than degrading everyone's install.

How it compares

The parts of ModelMRI are not new. Attention visualization, causal ablation, sparse autoencoders and agent tracing all have good tools already, and most of them do their one thing better than ModelMRI does. What is unusual here is the combination — model internals and agent traces, in one GUI, on your hardware, with no notebook in between.

Use TransformerLens if you are doing research and want precision. Use Langfuse or Phoenix if you are running an agent in production and need dashboards, retention and alerting. Reach for ModelMRI when you have a model on your machine that is doing something you do not understand, and you want to look at it now.

Questions people ask

What is ModelMRI?

An open-source, local-first tool for inspecting what a model is doing internally while it runs: attention per layer and head, which heads carry the prediction, which interpretable features fire, what changes when you turn one off — plus a recorder that makes an agent run inspectable step by step. pip install modelmri && modelmri serve, then open http://localhost:5900.

Does it work with GPT-4, Claude, or Gemini?

Not for internals, and no tool can — closed API models do not expose attention weights or activations to anyone outside the provider. ModelMRI needs weights it can run, so internals mean a local HuggingFace model.

The agent recorder is a different matter: it wraps whatever your agent calls, including hosted APIs, so you can record and inspect a run driven entirely by Claude or GPT-4. modelmri.record.instrument_anthropic() does it in one line.

Do I need a GPU?

No. CPU works — a 0.5B model streams in a couple of seconds. NVIDIA, AMD, Intel and Apple silicon are detected automatically when present, and if torch was installed as a CPU-only build the badge says so and prints the command to fix it.

Which models work?

Any causal LM transformers can load with eager attention, from the HuggingFace cache you already have, a plain folder on disk, or a search-and-download in the app. The ones with a recorded end-to-end result are GPT-2, Qwen3-0.6B, Qwen2.5-0.5B-Instruct, SmolLM2-360M-Instruct and Gemma-3-270m-it — the verified table lists what each one actually measured. Others should work and are not claimed to have been checked: Llama-3.2-1B is deliberately absent because the meta-llama repos are gated and returned 403 for the account used, and an untested model in a list headed "supported" is just a guess in bold. Ollama models work for text generation; internals need a HuggingFace model, and ModelMRI tells you that instead of quietly showing you nothing. Non-transformer models you trained yourself get a layer map with activation statistics.

Is any of my data sent anywhere?

No. There is no cloud, no account and no telemetry. Models are downloaded from HuggingFace if you ask for one you don't have; beyond that, nothing leaves the machine. A .mri file you choose to share contains tokens, attention and your note — never weights — and the browser viewer reads it client-side without uploading it.

What is a .mri file?

One ~54 KB file holding the tokens, the attention matrix, the generation and a note you wrote. It exists so you can send a colleague the finding without sending them 8 GB of weights or asking them to install anything — they open it at the viewer in a browser. modelmri open file.mri does the same locally in about 0.3s, with no torch and no GPU.

How is this different from TransformerLens or BertViz?

BertViz draws attention; ModelMRI also measures which heads matter by removing them. TransformerLens is a library you write code against and is more precise and more flexible than this; ModelMRI is a UI for the questions people ask most, and adds SAE steering, agent traces and robot policies in the same window. See How it compares.

Is it production-ready?

No, and the package says so — it is classified alpha. It is a debugging and research tool, not infrastructure. The measurements it reports are tested, but the API surface still moves between minor versions; see the changelog.

How do I cite it?

CITATION.cff is in the repository root — GitHub's "Cite this repository" button reads it. Please include the version, because the measured figures in this README belong to the release that produced them.

Contributing

Issues and pull requests are welcome. One rule runs the whole repository: don't ship a measurement you haven't verified. A visualization that looks plausible and is wrong is worse than none, because interpretability is exactly the domain where nobody has an independent way to notice.

• Contributing guide — setup, quality gates, and the three bugs that made that rule
• Code of conduct
• Security policy — trust model, credential handling, and what loading a model actually executes
• Support · Changelog

Built in public

Notes, mistakes, and what broke: modelmri.substack.com

MIT © Muhammad Mahad Azher