LlamaJEE.js View the research ↗
A SMALL-MODEL JEE PROJECT · BUILT IN PUBLIC

Work the problem.
Inspect every step.

A browser workspace for JEE Main and Advanced. Bring a question, follow the working, and check the result. Explore the models, mistakes, and experiments behind it.

Live demo: Qwen 0.8B9B browser export pendingModel details ↗

01 Question or screenshot02 Plan and solve03 Inspect the answer
THE WORKSPACE

Your question. The model’s working.

Exploratory output from a small browser model. Check every scientific step before relying on it.

01 / 03

Your question

Stays in this browser

Model workspace

Waiting
Q
Qwen3.5 0.8B · ONNXUnchanged browser demo model, separate from our 9B adapters
Checking this browser…

No account. No API key. No local server.
One-time download of roughly 850 MB. Requires WebGPU and enough device memory. We do not upload your question to our server; model files come from Hugging Face.

⌁

Your working will appear here.

The solver is a research preview. It can make confident mistakes, especially on Advanced multi-step questions.

THE MODEL SHELF · 23 OCTOBER 2026

One live demo. A preserved model history.

Choose with the full picture: the current browser demo and the trained 9B archive serve different purposes.

AVAILABLE HERE

0.8B browser demo

The model downloads from Hugging Face and runs on your device through WebGPU. Your questions and screenshots are processed in this browser. No sign-in, inference API or local server is needed.

ARCHIVED · NOT BROWSER-READY

Our 9B checkpoints

The unchanged 9B base and experimental adapters used MLX on a Mac. The adapter conversion to PEFT passed small native parity checks. A verified Transformers.js export is still missing.

View the actual adapter archive ↗
NEXT RELEASE GATE

Verify the converted model

Before enabling 9B here, we need browser-compatible weights, checks that the adapter is active, matched text and image outputs, and measured device memory. Publishing the page alone does not complete that work.

Read the export status ↗
THE METHOD

What the harness does

A browser-side structure around the model, designed for readable work rather than a bare answer.

02 / 03
01 / READ

Frame the problem

Subject, exam, expected answer type and possible thermodynamics sign conventions become explicit before generation.

02 / DECOMPOSE

Choose a route

The prompt asks for givens, a short subgoal plan, essential calculations and all requested branches.

03 / INSPECT

Show the limits

The page checks requested sections and runs a separate calculation for one simple kinematics pattern. Other mathematical steps remain unverified.

This first browser harness does not yet have a general symbolic checker or calculator. Those are research tasks, and the 9B MLX adapters cannot be loaded by Transformers.js as-is.

THE RECORD

The results, including the setbacks.

The 9B training work and this browser demo are separate experiments. The archive shows the actual checkpoints and what each taught us.

03 / 03
9B TRAINING ARCHIVE

No adapted checkpoint has been promoted as a better solver.

Our latest controlled studies changed training families and LoRA layer coverage. Some outputs improved, others regressed; none passed the full selection rule.

5experimental adapter snapshots
12questions in the latest development comparison
0checkpoints promoted as a general JEE gain
Three separate training studies: an early validation gain fell at later checkpoints, while the later stable run tied its unchanged base.
Checkpoint scores on separate development slices. Compare each line only with its own dashed unchanged-model baseline.
On the latest 12-question development set, the 16-layer adapter got one extra exact answer but fewer valid procedure points.
The latest 16-layer adapter reached 8/12 answers while procedure quality fell. These are small, reused development results.