3D Graph · Downselect · CLI · Contribute · API
There's no single place that answers: "What's the best model I can actually run on my hardware, for my use case, that my organization is allowed to use?"
ModelSpec answers that question by synthesizing data from HuggingFace, models.dev, LMArena, Artificial Analysis, and community research into a FalkorDB knowledge graph with a universal schema covering every model type.
Right now:
- 1,140+ model cards across 47 providers (OpenAI, Anthropic, Google, Meta, Mistral, NVIDIA, DeepSeek, Qwen, and 39 more)
- 8,400+ graph edges connecting models to providers, benchmarks, capabilities, platforms, licenses, and competitors
- 15 model types: LLM-Chat, LLM-Reasoning, LLM-Code, VLM, Embedding, Image Gen, Safety Classifier, Reward Model, Reranker, ASR, TTS, and more
- 750-field universal schema per model — one template for all model types
- 4-stage ranking engine with 9 use-case profiles and hardware-aware speed estimation
An interactive 3D force-directed visualization of the entire AI model landscape. Each model type has a unique 3moticon — a programmatic 3D mesh icon built with Three.js.
Features:
- 8 edge views: Provider, Lineage, Competition, Platforms, Benchmarks, Capabilities, Licensing, Tags
- 15 3moticon shapes: Speech bubble (chat), lightbulb (reasoning), code brackets (code), eye (VLM), rings (embedding), palette (image gen), shield (safety), star (reward), and more
- Click any node for a detail panel with properties and navigable relationships
- Search with instant results and camera focus
- Node size scales by parameter count, color by model type
Tell it what you need. It tells you which model to use.
Filter by:
- Use case: Coding, Reasoning, Chat, Embedding, Vision, Agentic, RAG, Safety
- Hosting: Local (self-hosted), Cloud Platform (AWS, Groq, Together), Provider API (OpenAI, Anthropic)
- Hardware: 14 presets from Raspberry Pi 5 to B200, with MIG partition support
- Constraints: Open weights, origin country, max cost, min context, parameter bounds
- Capabilities: Reasoning, tool use, vision
The ranking engine:
- Filters by hard constraints (memory fit, licensing, origin)
- Scores across benchmarks, capabilities, cost efficiency, and speed
- Ranks with tie-breaking by Arena ELO and parameter count
- Explains every score with per-dimension breakdowns
Select a GPU preset, optionally apply a MIG partition (1/2, 1/3, 1/4, 1/7), and the sliders auto-adjust memory, bandwidth, compute, and recommended quantization. The ranking engine estimates tokens/sec and filters out models that won't run at usable speed.
pip install -e ".[dev]"# Database overview
modelspec stats
# Search with filters
modelspec search --type llm-code --open-weights
# Model detail
modelspec info anthropic/claude-opus-4-6
# Side-by-side comparison
modelspec compare anthropic/claude-opus-4-6 openai/o3 google/gemini-2-5-pro
# Ranked recommendations
modelspec rank --use-case coding --top 10
# Hardware compatibility
modelspec hardware macbook_air_m4_24gb
# Find data gaps
modelspec gaps --top 20
# Auto-research from HuggingFace
modelspec research google/gemma-4-31b-it
# Validate all cards
modelspec validate
# Fork, branch, and PR your changes
modelspec contributeModelSpec is community-driven. Three ways to contribute research:
Visit /contribute in the web UI. Pick a model with data gaps, generate a research prompt, paste it into your favorite LLM (Claude, ChatGPT, Gemini), then paste the results back. The UI validates and submits a GitHub issue.
modelspec gaps --top 10 # Find what needs research
modelspec research <model_id> # Auto-fetch from HuggingFace
modelspec contribute # Open a PR with your changesModel cards live in models/{provider}/{model-slug}.md as YAML frontmatter + Markdown. Edit directly and submit a PR — GitHub Actions will validate the schema and report what changed.
FastAPI backend with full Swagger docs at /docs.
# Start the API
docker compose up -d # FalkorDB
pip install -e ".[dev]"
python scripts/ingest_all.py
uvicorn api.main:app --port 8000Endpoints:
| Method | Path | Description |
|---|---|---|
| GET | /api/v1/models |
List models (paginated, filterable) |
| GET | /api/v1/models/{id} |
Full model card as JSON |
| GET | /api/v1/graph?view=provider |
Graph data for 3D visualization |
| GET | /api/v1/search?q=gemma |
Search by name, provider, type |
| GET | /api/v1/stats |
Database statistics |
| POST | /api/v1/rank |
Ranking with use-case profiles |
| GET | /api/v1/views |
Available edge views |
| GET | /api/v1/node/{id} |
Node detail with relationships |
| GET | /graph |
3D knowledge graph UI |
| GET | /downselect |
Downselect wizard UI |
| GET | /contribute |
Community contribution UI |
/graph ─────┐
/downselect ┤──▶ FastAPI ──▶ FalkorDB Knowledge Graph
/contribute ┤ │ │
CLI ────────┘ │ 1,469 nodes
│ 8,472 edges
Ranking Engine │
(4-stage pipeline) │
│ Model Card Repo
│ (1,140 YAML files)
│ │
Scrapers ────────────┘
(HuggingFace, models.dev, AA)
Tech stack:
- Graph DB: FalkorDB (Redis-compatible, OpenCypher)
- Backend: Python 3.11+, FastAPI, Pydantic v2
- 3D Viz: Three.js via 3d-force-graph (zero build step)
- CLI: Typer + Rich
- Schema: 750-field Pydantic model with YAML serialization
Every model gets a universal YAML+Markdown card with 15 sections:
| Section | Fields | What it covers |
|---|---|---|
| Identity | 13 | Name, provider, type, status, release date |
| Architecture | 22 | Parameters, layers, attention, tokenizer |
| Lineage | 16 | Base model, training method, datasets |
| Licensing | 15 | License type, commercial use, origin country |
| Modalities | 60+ | Text, vision, audio, video, embeddings, reranking |
| Capabilities | 50+ | Coding, reasoning, tool use, languages, safety |
| Cost | 17 | Input/output pricing, cache, batch, fine-tuning |
| Availability | 83 | Platform-by-platform availability |
| Benchmarks | 50+ | Arena ELO, HumanEval, GPQA, MMLU, SWE-bench |
| Deployment | 30+ | Hardware profiles, runtimes, quantizations |
| Risk & Governance | 40+ | Bias evaluation, privacy, regulatory |
| Performance | 10 | Latency, throughput, quality/dollar |
| Adoption | 8 | Downloads, likes, community usage |
| Downselect | 14 | Compliance tags, eval status, custom scores |
| Sources | 12 | URLs and freshness tracking |
git clone https://github.com/turbobeest/modelspec.git
cd modelspec
# Start FalkorDB
docker compose up -d
# Install
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
# Ingest model cards into the graph
python scripts/ingest_all.py
# Start the API + web UI
uvicorn api.main:app --host 0.0.0.0 --port 8000
# Open in browser
# 3D Graph: http://localhost:8000/graph
# Downselect: http://localhost:8000/downselect
# Contribute: http://localhost:8000/contribute
# API Docs: http://localhost:8000/docsGenerate a logo for "ModelSpec" — a neural network knowledge graph platform. White transparent text on dark background. The "o" in "Model" is replaced by a stylized gear/cog made of interconnected neurons — nodes with synaptic connections forming a gear shape. Clean, technical, minimal. The neural-gear should subtly glow with a blue-indigo gradient (#3b82f6 to #6366f1). Monospace or geometric sans-serif font. Think: the intersection of machine learning and engineering precision.
MIT