Nimbus Model Lab

A learning and experimentation tool for real model checkpoints: follow supported tensor operations, examine weights, create GGUF variants and record inference results. Graph and Step-by-step share the same declared computation; 3D isolates an individual tensor.

Interactive architecture · Documentation

Install and run

Requires Python 3.12+, uv and Node.js 22.12+. Windows is the verified platform; macOS/Linux are unverified. Local checkpoint inspection works offline; Hub discovery/downloads require internet.

From the project root in PowerShell:

uv sync --locked
npm ci
npm ci --prefix apps/web
npm run build --prefix apps/web
.\scripts\start.ps1

Open http://127.0.0.1:8999.

Start

.\scripts\start.ps1

Status

.\scripts\status.ps1

Restart

.\scripts\restart.ps1

Stop

.\scripts\stop.ps1

Stop the server before updating Python dependencies on Windows. On macOS/Linux, use the same dependency/build commands, then uv run --no-sync llm-view; stop with Ctrl+C.

Native GGUF conversion, quantization and generation additionally need the pinned Windows CPU tools:

uv run --no-sync python scripts/setup_quantization.py

Workspaces

Section	Main actions
Library	Download pinned Hub files; open local/LM Studio checkpoints; organize and verify artifacts
Architecture	Graph, step equations, stored tensor tree, search, scope, tracing and SVG/PNG export
Weights	Full-plane heatmaps, exact cells, distributions, statistics and matching-baseline comparisons
Quantization	Exact tensor profiles, calibration, separate GGUF variants, error and corpus evaluation
Inference	Explicit generation, cached token steps, observed module captures and benchmarks
Compare	Two independently owned model workspaces with optional selection links
Experiments	Save, reproduce, compare and export recorded runs
System	Hardware history, provider capabilities, policies, appearance and runtime recovery

Support and limits

Metadata readers cover Safetensors, GGUF, restricted PyTorch, TensorFlow, Flax and ONNX. Numerical inspection depends on encoding and reader limits.

Model family	Graph / Step-by-step	Stored tensor tree / individual 3D tensor
GPT-2 (gpt2, real checkpoints)	Supported	Supported
Llama (llama)	Supported	Supported
Mistral (mistral)	Supported	Supported
Qwen2 (qwen2)	Supported	Supported
Qwen3 (qwen3)	Supported	Supported
Qwen3.5 text (qwen3_5, qwen3_5_text)	Supported, full/linear attention	Supported
Mixtral (mixtral)	Supported, MoE	Supported
Qwen2 MoE (qwen2_moe)	Supported, MoE	Supported
Qwen3 MoE (qwen3_moe)	Supported, MoE	Supported
Embedding Gemma 2 / Gemma 4	No validated adapter	Available when container metadata is readable
Nemotron-H MoE	No validated adapter	Available when container metadata is readable
Other families; vision/audio/MTP or separate LoRA components	No validated adapter	Available when container metadata is readable

Graph support requires valid configuration and compatible bindings. Custom auto_map, cross-attention and unrecognized attention schedules are rejected. TensorFlow/Flax layouts have inventory support without family graph bindings; ONNX imports its exported top-level graph separately. Tensor inspection is based on the container, not the model family. Heatmaps and statistics additionally require a supported numeric encoding.

• Weights alone do not establish computation. Arbitrary model-code tracing is not implemented.
• Inference requires an explicit action and a supported local runtime. Native GGUF has no intermediate activation captures.
• Restart expires live sessions; recorded results remain local.
• The service is for a local owner, binds loopback and has no authentication. Keep credentials, checkpoints and runtime data out of Git.

See format and adapter limits and security.

Development

.\scripts\check.ps1
node docs/architecture/check.mjs
uv build

The checks cover types, tests, lint, formatting and the frontend build. Build the frontend before packaging; the wheel includes its output. Frontend development uses npm run dev --prefix apps/web, with /api proxied to port 8999 (API_PROXY overrides it). LLM_VIEW_HOST and LLM_VIEW_PORT configure the Python server.

Architecture on GitHub Pages

In GitHub Settings → Pages, choose Deploy from a branch, main, /docs, then Save. docs/.nojekyll serves the viewer as static files. The architecture link becomes available after publishing and enabling Pages. No custom workflow is required.

License and credits

Apache 2.0: reuse, modification and commercial use are allowed. Redistribution must retain required notices and identify modified files; see NOTICE. The learning purpose is an intended use, not a license restriction. Three.js r160 retains its separate MIT notice. Tensor presentation draws on GenAI Tensor Tree. Downloaded models and native tools retain their own licenses.

Citation

If you use Model-Lab in your work, please cite this repository.

@misc{mugilnimbus2026modellab,
  author = {{mugilnimbus}},
  title  = {Nimbus Model Lab},
  year   = {2026},
  url    = {https://github.com/mugilnimbus/Model-Lab}
}