- Serverless: All of Synth’s algorithms are designed to work with your code with minimal affordances. If you can wrap your code in 3 simple HTTP routes, swap out your chat completions / responses url, and spin up a free Cloudflare tunnel, you can do GEPA, RL, and SFT.
- Composable: Models trained and prompts calibrated in one session can be used in any other. Programmatically pipelining multiple training methods requires only a few SDK calls.
- Human-usable: All methods provide the user with rich observability, cancellability, and where appropriate, interventions.
Highlights
- 🚀 Serverless Training: Train across SFT, RL, and prompt optimization by standing up a single Cloudflare FastAPI wrapper around your code. No production code churn.
- ⚡️ High Performance: Parallelize training and achieve 80% GPU utilization via PipelineRL
- 🗂️ Multi-Experiment Management: Train prompts and models across multiple experiments with unified tracking
- 🛠️ Local Development: Spin up experiment queues and datastores locally for dev work
- 🔩 Serverless Training: Run training via CLI or programmatically without infrastructure management
- 🏢 Enterprise Scale: Scales GPU-based model training to 64 H100s seamlessly
- 💾 GEPA-Calibrated Judges: Use fast, accurate rubric scoring with GEPA-calibrated judges
- 🖥️ Language Agnostic: Supports HTTP-based training across all programming languages
- 🤖 CLI Integration: CLI utilities tuned for use with Claude Code, Codex, and Opencode
- 📊 Comprehensive Tracking: Full observability with traces, metrics, and event logs
Quick Start
Get started with Synth AI in minutes:GEPA (Recommended)
Evolutionary prompt optimization - fastest way to get started
MIPRO
Meta-learning prompt optimization with TPE
Reinforcement Learning
Train model weights with GSPO policy gradient
Supervised Fine-Tuning
Fine-tune on demonstration data
Installation
Install Synth AI with uv:What’s Next?
- First Steps - Get GEPA running on Banking77 in 5 minutes
- Training Methods - GEPA, MIPRO, GSPO, SFT reference
- CLI Reference - Command-line interface guide
- SDK Reference - Programmatic access
- Task Apps - Build custom training environments
- llms.txt - Instructions for coding agents (Claude Code, Codex)