ExocoreV1
A from-scratch AI system built around ExocoreV1 — a powerful 751M-parameter language model, fully rebranded and independent of any third-party naming.
HuggingFace: https://huggingface.co/Exocore/ExocoreV1
Project Overview
ExocoreV1 is a full-stack AI chat and training platform:
- Chat interface — Talk to ExocoreV1 with Fast, Thinking, and Deep Think modes
- Research tools — Web search (DuckDuckGo) + multi-URL reader integrated into the chat pipeline
- Training page — Train new models from text, with live Chart.js loss curve and architecture tree visualization
- Auto-train — Gemini-powered automatic data generation + training
- Auto-load — Exocore.pt loads automatically on startup (no button click needed)
Architecture
- Backend: Python + Flask (
main.py) - Model engine:
exocore_model.py— ExocoreLM (28-layer transformer, 751M params, GQA, RoPE, SwiGLU) - Training:
trainer.py— MiniGPT character-level trainer + ExocoreV1 inference - Gemini client:
gemini_client.py - Templates:
templates/index.html,templates/train.html
Key Files
| File | Purpose |
|---|---|
main.py |
Flask server, all API routes |
exocore_model.py |
ExocoreLM architecture (Qwen3-compatible, ExocoreV1 branded) |
trainer.py |
Training loop + MiniGPT + ExocoreV1 inference |
gemini_client.py |
Gemini API integration |
models/trained/Exocore.pt |
ExocoreV1 model weights (751M params, PROTECTED) |
ExocoreV1 Model Architecture
| Component | Value |
|---|---|
| Parameters | 751.6M |
| Layers | 28 |
| Hidden size | 1024 |
| Attention heads | 16Q / 8 KV |
| Head dim | 128 |
| FFN intermediate | 3072 (SwiGLU) |
| Vocab size | 151,936 |
| Max context | 40,960 tokens |
| RoPE theta | 1,000,000 |
| Precision | bfloat16 |
API Endpoints
| Endpoint | Method | Description |
|---|---|---|
/api/models |
GET | List all available models |
/api/load |
POST | Load a model |
/api/chat |
POST | Stream chat (SSE) |
/api/clear |
POST | Clear chat history |
/api/status |
GET | Server/model status |
/api/autoload-status |
GET | Auto-load progress |
/api/model-details |
GET | Current model architecture info |
/api/search |
POST | DuckDuckGo web search |
/api/read-link |
POST | Fetch and parse URL content (multi-link) |
/api/train/start |
POST | Start training |
/api/train/status |
GET | Training status + loss history |
/api/train/stop |
POST | Stop training |
/api/train/models |
GET | List trained models |
/api/train/model-info/<name> |
GET | Model architecture info |
/api/gemini/* |
POST/GET | Gemini integration |
Chat Modes
| Mode | Description |
|---|---|
| Fast | Direct answer, no thinking |
| Thinking | Single-pass reasoning before answering |
| Deep Think | Two-pass multi-step reasoning (Pass 1 + Pass 2 think blocks) |
Research Features
- Web search: DuckDuckGo HTML scraping, returns title + snippet + URL
- Read Page: Fetches any URL, strips nav/scripts, returns clean text (up to 4000 chars)
- Multi-link reading: Read up to 4 URLs simultaneously and merge into research context
- Context injection: Research context is automatically injected into the model prompt
User Preferences
- No system prompt exposed in UI — ExocoreV1 identity is baked into the model prompt automatically
- Exocore.pt is permanently protected from deletion
- No "qwen" branding anywhere in the UI or code
- Model auto-loads on startup
- No system prompt field in UI — identity lives inside the model