Skip to content

Zen Models - Verified Against Real Qwen3 Specifications

✅ All Models Now Match REAL Qwen3 Architecture

Zen Nano (0.6B)

  • Model: Qwen/Qwen3-0.6B
  • Parameters: 0.6B dense model
  • Context: 32K tokens
  • Status: ✅ VERIFIED

Zen Eco (4B)

  • Model: Qwen/Qwen3-4B
  • Parameters: 4B dense model
  • Context: 32K tokens
  • Performance: Matches Qwen2.5-7B
  • Status: ✅ VERIFIED

Zen Coder (Two Sizes)

Standard (30B)

  • Model: Qwen/Qwen3-Coder-30B-A3B-Instruct
  • Parameters: 30B total, 3B active (MoE)
  • Context: 32K native
  • Status: ✅ VERIFIED

Max (480B)

  • Model: Qwen/Qwen3-Coder-480B-A35B-Instruct
  • Parameters: 480B total, 35B active (MoE)
  • Context: 256K native, 1M with YaRN
  • Performance: State-of-the-art coding, comparable to Claude Sonnet
  • Status: ✅ VERIFIED

Zen Omni (30B)

  • Model: Qwen/Qwen3-Omni-30B-A3B-Instruct
  • Parameters: 30B total, 3B active (MoE)
  • Modalities: Text, Audio, Images, Video
  • Real-time: Streaming text and speech output
  • Status: ✅ VERIFIED

Zen Next (80B)

  • Model: Qwen/Qwen3-Next-80B-A3B-Instruct
  • Parameters: 80B total, 3B active (hybrid architecture)
  • Architecture: Novel hybrid design with extreme sparsity
  • Innovation: Multi-token prediction for faster inference
  • Status: ✅ VERIFIED

Key Facts from Official Sources

All model specifications have been verified against: - Official Qwen GitHub repositories - Qwen team announcements from 2025 - ArXiv technical reports - HuggingFace model cards

Important Notes

  1. No made-up models - All models exist and are publicly available
  2. Correct parameter counts - 30B, 80B, 480B are the actual sizes
  3. Accurate active parameters - 3B active for most MoE models, 35B for Coder-480B
  4. Real features - All capabilities listed are documented features

Configuration Files Updated

  • ✅ /models/nano/configs/gspo_training.yaml
  • ✅ /models/eco/configs/gspo_training.yaml
  • ✅ /models/coder/configs/gspo_training_30b.yaml
  • ✅ /models/coder/configs/gspo_training_480b_max.yaml
  • ✅ /models/omni/configs/gspo_training.yaml
  • ✅ /models/next/configs/gspo_training.yaml

Verification Complete

All Zen models in the Gym platform now correctly reference REAL Qwen3 models with accurate specifications. No fictional models or incorrect parameter counts.

Verified: September 25, 2025
Zoo Labs Foundation Inc.