mirror of
https://github.com/zenlm/gym.git
synced 2026-11-04 17:11:44 +00:00
Major updates: - Implemented BitDelta 1-bit quantization (10-25× memory reduction) - Added DeltaQuant flexible quantization (INT2/4/8, Binary, Ternary) - Created DeltaSoup for Byzantine-robust community aggregation - Configured accurate Qwen3 models for Zen family (Nano, Eco, Coder, Omni, Next) - Added GRPO/GSPO training algorithms with memory optimizations - Simplified CLI from gym-cli to gym - Updated WebUI with Zoo Labs black monochromatic theme - Integrated Unsloth optimizations for 2-3× speedup - Added comprehensive test suite for quantization methods Zoo Labs Foundation Inc. - Copyright 2025
1.1 KiB
1.1 KiB
Zen Next Model (32B MoE)
Advanced reasoning and complex problem-solving
Overview
Zen Next represents cutting-edge capabilities for:
- Advanced reasoning tasks
- Complex problem-solving
- Research applications
- Enterprise solutions
- High-accuracy requirements
Model Details
- Base Model: Qwen/Qwen3-Next-32B-MoE-Instruct
- Parameters: 32B (8B active per token via MoE)
- Context Length: 4096 tokens
- Memory Required: ~48GB VRAM
Training
Quick Start
# Train with GRPO algorithm
gym train models/next/configs/gspo_training.yaml
# DeepSpeed integration
gym train models/next/configs/gspo_training.yaml \
--deepspeed configs/deepspeed/ds_z3_config.json
Configuration
- Algorithm: GSPO (MoE-optimized for scale)
- LoRA Rank: 64
- Learning Rate: 1e-5
- Batch Size: 1
Performance
- MMLU: 81.2
- HumanEval: 78.9
- MT-Bench: 8.9
- Speed: 30 tokens/sec
Use Cases
✓ Research tasks ✓ Complex analysis ✓ Strategic planning ✓ Scientific computing ✓ Advanced AI agents
Part of the Zen Model Family by Zoo Labs Foundation