Spaces:
Running
title: Matrix.Corp
emoji: 🐦🔥
colorFrom: indigo
colorTo: pink
sdk: static
pinned: true
Matrix.Corp
Independent AI research organization building specialized models, frontier agentic systems, and new intelligence paradigms.
HuggingFace: Matrix-Corp · Founded by: Zandy-Wandy · GitHub: zapgaming
Status Legend
| Badge | Meaning |
|---|---|
| 🟢 Released | Weights available, ready to use |
| 🟡 Preview | Architecture published, training planned |
| 🔴 Planned | Design complete, not yet built |
| 🩵 Long-Term | Vision defined, major research ahead |
| 🟣 Closed | Proprietary weights |
| ⬛ Deprecated | Cancelled or superseded |
Models
🌌 Zenith — Reasoning + Emotional Intelligence
Status: 🟡 Preview · Target: Tenstorrent Blackhole p300a
Transformer models with a built-in EQ Engine — a dedicated emotional intelligence layer that sits alongside the reasoning stack. Ring Attention (32K), MoE (12 experts top-2), Ollama + vLLM compatible.
| Model | Params | Base | Link |
|---|---|---|---|
| Zenith-7B-V1 | 7B | Qwen2.5-Coder-7B | → |
| Zenith-28B-V1 | 28B | Qwen3.5-27B (Opus 4.6 distilled) | → |
| Zenith-32B-V1 | 32B | DeepSeek-R1-Distill-Qwen-32B | → |
| Zenith-70B-V1 | 70B | DeepSeek-R1-Distill-Llama-70B | → |
🔬 Vortex Scientific — Deep Science Reasoning
Status: 🟡 Preview · Target: MacBook M2/M3 + Nvidia 4060
Built from scratch — no base model. Custom 50K science tokenizer. Hybrid SSM+Attention architecture with four domain-specific modules: Equation/LaTeX, Numerical, Citation, and Molecular/Periodic Table.
🌿 Touch Grass — Music AI
Status: 🟡 Preview · Target: Any hardware
LoRA fine-tune on Qwen3.5 built for musicians. Tab & Chord Module, Music Theory Engine, Ear Training, EQ Adapter (4 emotional modes), Songwriting Module.
| Model | Params | Base | Link |
|---|---|---|---|
| TouchGrass-3B | 3B | Qwen3.5-3B-Instruct | → |
| TouchGrass-7B | 7B | Qwen3.5-7B-Instruct | → |
🌐 Matrix Lattice — Frontier Agentic MoE
Status: 🟢 Released · 🟣 Closed Source · Target: 4–32× H100 / Tenstorrent p300a
Shipped. Our largest and most capable system. Frontier-scale mixture-of-experts with 17 custom intelligence modules including: EQ Engine V2, Multi-Agent Coordination Layer (MACL), Hierarchical Context Compression Engine (HCCE), Causal Reasoning Graph, Long-Horizon Task Planner, Confidence Calibration Head, Safety Reasoning Module (SRM), and more. 1M token context across all tiers.
| Model | Total Params | Active Params | Experts | Context | Link |
|---|---|---|---|---|---|
| Lattice-120B | 120B | ~22B | 64 top-4 | 1M | → |
| Lattice-430B | 430B | ~38B | 128 top-4 | 1M | → |
| Lattice-671B | 671B | ~47B | 256 top-4 | 1M | → |
🩸 Matrix ECHO — Living Error Memory
Status: 🔴 Build In Progress · 🟢 Open Source · Language: Rust
The model that remembers how it was wrong.
ECHO is a 27B coding-focused LLM built on Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled, running fully in Rust via HuggingFace candle. Every correction it receives crystallizes into a Scar — a typed, weighted memory object stored in a live petgraph lattice.
Before every response, ECHO scans its Scar lattice for similar past mistakes. The more it's corrected, the harder it is to fool. Mistakes are not erased — they become assets.
Core loop:
prompt → pre-scan Scar lattice → inject caution context → generate → correction → new Scar forms
Scar types: Factual · Logical · Contextual · Hallucination · Overconfidence
Domain Weakness Map — ECHO tracks which topics it's systematically weak in and suppresses confidence automatically in high-risk domains.
OpenAI-compatible API — drop-in via POST /v1/chat/completions. Corrections via POST /v1/echo/correct.
| Model | Params | Base | Language |
|---|---|---|---|
| ECHO-27B-V1 | 27B | Qwen3.5-27B (Opus 4.6 distilled) | Rust + candle |
🎨 Matrix Voxel — 3D Generation
Status: 🔴 Planned · Target: A100 40GB
Flow-matching DiT backbone (~2.3B) with task-specific decoder heads. Generates 3D meshes, environments, printable models, and NeRF/Gaussian Splatting outputs.
| Model | Task | Outputs | License |
|---|---|---|---|
| Voxel Atlas | World/environment gen | .vox, .obj, .usd | 🟢 Open |
| Voxel Forge | 3D mesh & assets | .obj, .glb, .fbx, .usdz | 🟢 Open |
| Voxel Cast | 3D printable | .stl, .step, .3mf | 🟢 Open |
| Voxel Lens | NeRF / Gaussian Splatting | .ply (3DGS) | 🟢 Open |
| Voxel Prime | Unified all-in-one | All formats | 🟣 Closed |
🔷 Matrix Vexa — Crystalline Intelligence Substrate
Status: 🔴 Paused · 🟢 Open Source
Vexa is not a model. It is a new intelligence paradigm — a living lattice of Glyphs (structured meaning objects) that grows through Crystallization instead of training. 10 minutes on any CPU. No GPU required. Knowledge never goes stale — three background threads continuously update from the web, interactions, and decay.
Full paradigm definition and build prompt complete. Build paused, will resume.
⬛ Kairiq — Critical Moment Intelligence Module
Status: ⬛ Deprecated
A Lume-native intelligence amplifier module designed to wrap Matrix models. Deprecated — the custom Lume language runtime exceeded practical build complexity. Core ideas (pre-scan, confidence suppression, domain routing) absorbed into ECHO.
Paradigms
| Name | Type | Status |
|---|---|---|
| Crystalline Intelligence (Vexa) | Non-neural knowledge substrate | 🔴 Paused |
| Living Error Memory (ECHO) | Scar-based mistake crystallization | 🔴 Build In Progress |
| Ferric Attention | Ownership-typed attention mechanism | 🩵 Research concept |
Reserved Names
These names are allocated to specific projects. Not available for other uses.
| Name | Allocated To |
|---|---|
| Vexa | Crystalline Intelligence Substrate |
| ECHO | Living Error Memory LLM |
| Axiom | Future extreme reasoning model (planned) |
| Lume | Declarative-relational language for Vexa |
Licensing
| Model Family | License |
|---|---|
| Zenith | Apache 2.0 |
| Vortex | Apache 2.0 |
| Touch Grass | Apache 2.0 |
| Matrix Lattice | Proprietary |
| Matrix ECHO | Apache 2.0 |
| Matrix Voxel (open tiers) | Apache 2.0 |
| Matrix Voxel Prime | Proprietary |
| Vexa | Apache 2.0 |
Matrix.Corp — building intelligence that knows its own limits.