---
language:
- en
license: mit
library_name: transformers
tags:
- multimodal
- moe
- text-to-image
- image editing
- image to video
- text-to-video
- video editing
- text-to-speech
- speech-to-text
- speech-to-speech
- image-to-text
- video-to-text
- agentic
- tool-use
- flow-matching
- 3d-rope
- titok
- vidtok
- dual-stream-attention
- zero-shot-voice-cloning
- bigvgan
- snake-activation
- multi-receptive-field-fusion
pipeline_tag: any-to-any
inference: false
datasets:
# === Code & Programming ===
- m-a-p/Code-Feedback
- iamtarun/python_code_instructions_18k_alpaca
- codeparrot/codeparrot-clean
- bigcode/humanevalpack
- loubnabnl/github-jupyter-code-to-text
- saurabh5/rlvr-code-data-Swift
- finbarr/rlvr-code-data-swift-code-edit
- ExAi/Code-Golang-QA-2k
- smcleod/golang-coder
# === Conversation & Agentic ===
- databricks/databricks-dolly-15k
- OpenAssistant/oasst1
- HuggingFaceH4/no_robots
- Open-Orca/OpenOrca
- abhi227070/conversation-to-summarization-dataset
- allenai/WildChat-1M
- THUDM/AgentInstruct
- glaiveai/glaive-code-assistant-v2
- stingning/ultrachat
- RyokoAI/ShareGPT52K
- AlicanKiraz0/Agentic-Chain-of-Thought-Coding-SFT-Dataset
# === Tool Use ===
- Locutusque/function-calling-chatml
- driaforall/pythonic-function-calling
- argilla/Synth-APIGen-v0.1
- interstellarninja/tool-calls-singleturn
- interstellarninja/tool-calls-multiturn
# === Vision (Image & Video) ===
- Naveengo/flickr8k
- ybelkada/football-dataset
- jmhessel/newyorker_caption_contest
- derek-thomas/ScienceQA
- HuggingFaceM4/WebSight
- lmms-lab/Video-MME
- MBZUAI/VideoInstruct-100K
# === Generation (Prompts & Media) ===
- Gustavosta/Stable-Diffusion-Prompts
- FredZhang7/stable-diffusion-prompts-2.47M
- succinctly/midjourney-prompts
- osunlp/MagicBrush
- timbrooks/instructpix2pix-clip-filtered
- Rapidata/sora-video-generation-physics-likert-scoring
- Rapidata/sora-video-generation-style-likert-scoring
- Rapidata/sora-video-generation-alignment-likert-scoring
- Rapidata/text-2-video-human-preferences
- Rapidata/text-2-video-human-preferences-sora-2
- TempoFunk/webvid-10M
- multimodalart/panda-70m
- nkp37/OpenVid-1M
- WenhaoWang/VidProM
- WenhaoWang/TIP-I2V
- jovianzm/img2vid-pexels-350k
- TencentARC/MiraData
- APRIL-AIGC/UltraVideo
- Mutonix/Vript
- Rapidata/image-to-video-human-preference-seedance-1-pro
# === Audio ===
- openslr/librispeech_asr
- blabble-io/libritts_r
- parler-tts/mls_eng_10k
- MikhailT/hifi-tts
# === File Ops ===
- renjiepi/medium_20000-file_operations_n100k1
---

# 🚀 Xoron-Dev: State-of-the-Art Multimodal MoE

<div align="center">

![Xoron-Dev Logo](https://img.shields.io/badge/Xoron--Dev-MultiMoE-blue?style=for-the-badge&logo=pytorch)
![License](https://img.shields.io/badge/License-MIT-green?style=for-the-badge)
![Params](https://img.shields.io/badge/Parameters-4B_MoE-yellow?style=for-the-badge)
![Context](https://img.shields.io/badge/Context-128K-red?style=for-the-badge)
![Version](https://img.shields.io/badge/Version-2.2-purple?style=for-the-badge)

</div>

<p align="center">
  <img src="assets/IMG_2970.png" alt="Training-Stage" width="200">
</p>


# ![Xoron-Dev Logo](assets/IMG_2925.PNG)
**Xoron-Dev**


# ✨ Xoron-Dev: The Elite SOTA Omni-Modal Intelligence

**Xoron-Dev** is the definitive open-source architecture for **Omni-Modal Artificial Intelligence**. Unlike legacy models that treat vision and audio as plugins, Xoron-Dev is designed for native, high-fidelity perception across every major sensory dimension.

---

## 🌟 Why Xoron-Dev?

Xoron-Dev represents a massive leap in multimodal reasoning, combining cutting-edge Sparse MoE architecture with a refined sensory stack.

### 1. 👁️ SOTA Vision (SigLIP-2 & TiTok)
Xoron-Dev exclusively uses **SigLIP-2** for superior zero-shot performance and semantic alignment.
- **TiTok 1D VAE:** Images are compressed into **256 ultra-dense tokens**, allowing Xoron to "see" high-resolution scenes with unprecedented efficiency.
- **2D-RoPE:** Integrated positional embeddings that maintain spatial relationships regardless of aspect ratio.

### 2. 🎬 Native Video Intelligence (VidTok)
Our custom **VidTok** encoder uses **3D Volumetric Compression** to ingest up to **32 frames** of high-definition video natively. Xoron doesn't just see a sequence of images—it understands motion, causality, and temporal context.

### 3. 🎙️ Raw PCM Audio (Conformer + BigVGAN)
Xoron-Dev processes **Raw 16kHz PCM Audio** directly. No Mel Spectrograms, no lossy Fourier transforms.
- **Micro-Latency S2S:** True Speech-to-Speech interactions (<200ms) for natural, fluid conversations.
- **Zero-Shot Voice Cloning:** Instantly clone any voice from a 5-second sample for high-fidelity personalized output.

### 🧠 The Brain: Aux-Lossless MoE & 128K Ring Attention
A sophisticated **Mixture of Experts** (MoE) backbone that dynamically routes the logic of every token through specialized hardware-aware sub-networks.

#### 🏗️ Deep Expert Hierarchy
Unlike standard MoE models with uniform experts, Xoron-Dev implements a specialized **Deep Expert** system.
- **Expert Pool:** 16 Experts Total (8 Standard + 8 Deep).
- **Variable Logical Depth:** Deep Experts possess internal depths scaling from **2 up to 9** layers.
- **Expert Penalty Routing:** A soft utilization penalty ($Cost \propto Depth$) ensures that the model only invokes deeper computation for tasks requiring maximum logical precision, maintaining high inference throughput for simpler tokens.

#### ⚡ Reasoning Acceleration: Fast Ponder
Xoron-Dev features a dedicated **FastPonderBlock** for near-instant latent deliberation.
- **Attention-Free Reasoning:** By bypassing the $O(N^2)$ Self-Attention stack during thought loops, the Depth-3 reasoning block propagates logic at **120+ thoughts/sec**.
- **Dynamic Halting:** A learned `halt_head` monitors latent entropy. Once the model reaches a decision (entropy threshold < 0.2), it breaks the ponder loop and returns to token decoding, reducing unnecessary FLOPs by up to 90%.

#### 🔘 Infinite Context
Using **Ring Attention**, Xoron-Dev can analyze books, hour-long videos, or massive codebases with native **128K context window** support.

---

## 🚀 Get Started with Xorfice

The easiest way to experience Xoron-Dev is via the `xorfice` engine—the SOTA orchestrator for multimodal deployment.

### Installation
```bash
pip install xorfice
```

### High-Fidelity Interaction
```python
from xorfice import XoronEngine

# The engine automatically handles weights and optimizations
# Correct model slug: Backup-bdg/Xoron-Dev-MultiMoe
engine = XoronEngine(model_path="Backup-bdg/Xoron-Dev-MultiMoe")

# Start an omni-modal conversation
response = engine.generate(
    prompt="Who is this person and what are they doing?",
    images="https://example.com/interview.jpg",
    videos="https://example.com/interview.mp4"
)
print(response["text"])
```

---

## 📈 SOTA Benchmarks & Features

| Feature | Xoron-Dev |
| :--- | :--- |
| **Vision Backbone** | SigLIP-2 |
| **Video Compression** | VidTok 3D |
| **Audio Ingestion** | Raw PCM |
| **Inference Efficiency** | Sparse MoE (5B) |
| **Context Window** | 128K (Ring) |

---

## 🎨 Creative Generation
Fully integrated with **MobileDiffusion**, Xoron-Dev doesn't just understand—it creates.
- **Text-to-Video (T2V)**
- **Image-to-Video (I2V)**
- **Text-to-Image (T2I)**
- **Image-to-Image (I2I)**
- **Video-to-Video (V2V)**

---

### Join the Revolution
Xoron-Dev is more than a model—it's a vision for the future of AI. Build your own multimodal agent today.

*Powered by [Xoron-Dev Team](https://xoron.dev)*