Open Foundation Models for Agentic AI
30+ models from 0.6B to 1T parameters across language, vision, audio, video, and 3D
Zen LM provides production-ready AI models for agentic coding, multimodal understanding, and creative generation. Our flagship Zen Coder models are trained on 8.47 billion tokens of real-world programming sessions, delivering state-of-the-art performance on agentic programming tasks.
Zen Coder - Agentic Coding Models
Fine-tuned on 8.47B tokens of real programming sessions
| Model | Size | Base | VRAM | Context | Status |
|---|---|---|---|---|---|
| Zen Coder 4B | 4B | Qwen3-4B-Instruct | 8 GB | 32K | Trained |
| Zen Coder 24B | 24B | Devstral Small 2 | 24 GB | 256K | Trained |
| Zen Coder 123B | 123B | Devstral 2 | 128 GB | 256K | Training |
| Zen Coder Max | 358B | GLM-4.7 (MoE) | 180 GB | 200K | Planned |
| Zen Coder Ultra | 1T | Kimi K2 (MoE) | 256 GB | 128K | Planned |
Real Agentic Data
Trained on actual agentic debug sessions - not synthetic data. Real debugging workflows, multi-file refactoring, and tool use patterns.
Production Code
15 years of professional development across AI, Web3, cryptography, and modern software engineering from 1,452 repositories.
Open Training
Use zen-trainer to fine-tune on your own data. Supports MLX (Apple Silicon), Unsloth, and DeepSpeed.
Complete AI Model Ecosystem
Language Models
6 core models from 0.6B to 32B. zen-nano for edge, zen-eco for efficiency, zen-omni for multimodal, zen-next for frontier reasoning.
Zen Coder
5 coding models from 4B to 1T trained on 8.47B tokens of agentic programming data. State-of-the-art on tool use and multi-step coding.
Vision & Multimodal
zen-vl for vision-language, zen-designer for visual understanding, zen-artist for image generation, zen-omni for unified multimodal.
Video & 3D
zen-director for video generation, zen-video for high-quality synthesis, zen-3d for 3D assets, zen-world for world simulation.
Audio
zen-musician for music generation, zen-foley for sound effects, zen-scribe for transcription, zen-dub for voice dubbing.
Specialized
zen-guard for safety, zen-embedding for vectors, zen-reranker for search, zen-translator for translation, zen-agent for tool use.
Zen Agentic Dataset
8.47 Billion Tokens of Real-World Agentic Programming
Tokens
Total training tokens across all data sources
Samples
Training samples with conversation context
Repositories
Open source and private codebases
History
Years of development history (2010-2025)
Get Started
zen-trainer
Fine-tune models on your own data
pip install zen-trainer