mirror of
https://github.com/zenlm/gym.git
synced 2026-11-04 17:11:44 +00:00
Zen Omni Model (14B MoE)
Multimodal capabilities for diverse applications
Overview
Zen Omni supports multimodal understanding across:
- Vision-language tasks
- Audio processing
- Cross-modal reasoning
- Creative applications
- Multimodal analysis
Model Details
- Base Model: Qwen/Qwen3-Omni-14B-MoE-Instruct
- Parameters: 14B (4B active per token via MoE)
- Context Length: 2048 tokens
- Memory Required: ~24GB VRAM
Training
Quick Start
# Train with GSPO algorithm
gym train models/omni/configs/gspo_training.yaml
# Enable vision tower
gym train models/omni/configs/gspo_training.yaml \
--freeze_vision_tower false
Configuration
- Algorithm: GSPO (MoE-stabilized for multimodal)
- LoRA Rank: 32
- Learning Rate: 1.5e-5
- Batch Size: 1
Performance
- MMLU: 74.6
- HumanEval: 65.2
- MT-Bench: 8.3
- Speed: 45 tokens/sec
Use Cases
✓ Image captioning ✓ Visual Q&A ✓ Audio transcription ✓ Multimodal chat ✓ Creative tools
Part of the Zen Model Family by Zoo Labs Foundation