Skip to main content
Synth is a post-training platform for agents. You ship a task app (a small FastAPI service that exposes your environment), collect traces through structured rollouts, and then run either supervised fine-tuning (SFT) or reinforcement learning (RL) to improve the policy. Every flow shares the same core pieces:
  1. Task app – describes actions, observations, rubrics, and tracing hooks.
  2. Rollouts / eval – run the agent to gather data, judge results, and benchmark policies.
  3. Training job – submit an SFT or RL job with the Synth CLI, monitor progress, and retrieve the updated model.
Agents (humans or other services) can read these pages top-to-bottom and reproduce the entire workflow.

​
How the docs are organized

  • Quickstart – install the CLI, pair it with your dashboard session, and walk through the math demo.
  • SFT docs – gather or filter JSONL datasets, configure hyperparameters, and launch fine-tuning jobs.
  • RL docs – configure trainer topologies, deploy task apps to Modal, and run online RL.
  • Rollouts – the uvx synth-ai eval command covers both evaluation and data collection.
  • Reference material such as pricing and models lives in their dedicated sections.

​
Model families (Synth hosted)

ModelSizesNotes
Qwen 30.6B, 1.7B, 4B, 8B, 14B, 32BBest starting point, dense thinking models that support tool-calling
Qwen 3 (Advanced)4B-2507, 30B-A3B, 235B-A22B*, 480B-A35B*Unique Instruct & Thinking variants with MoE support
Qwen 3 VL2B, 4B, 8B, 30B-A3B, 32B, 235B-A22B*Multimodal (vision + language) family with Instruct & Thinking variants
Qwen 3 Coder30B-A3B, 480B-A35B*Specialized for code generation (Instruct only)
* 235B and 480B models must be sharded across multiple GPUs for inference and training.