Up Next
- Enhanced multi-stage pipeline optimization
- Expanded model provider support
- Improved developer experience tools
- More production-ready examples and cookbooks
Changelog
- synth-large model now available in Research Agent harness for MIPRO optimization
- Improved MIPRO and GEPA visualizations in the dashboard
- Research Agents: Automated prompt optimization in sandboxed environments
- Language-Agnostic: OpenAPI contracts enable Task Apps in any language
- Multi-interface access: Web Dashboard, Python SDK, and CLI for Research Agents
- Complete polyglot examples in Rust, Go, TypeScript, and Zig
- In-Process Task Apps: Complete MIPRO and GEPA optimization with automatic task app and tunnel management
- Gemini 3 Model Support: Google’s latest Gemini 3 models now supported
- Production-Ready Examples: Complete pipeline examples for GEPA and MIPRO
- SDK Release 0.2.25.dev1: Latest package updates
- Artifacts CLI: Unified command suite for managing all Synth AI artifacts
- Experiment Queue System: Redis-based queue with Celery workers
- Session-Based Pricing: Comprehensive usage tracking and budget enforcement
- Cloudflare Tunnel Support: Quick tunnels and managed tunnels for task app deployment
- Task App Discovery: Multi-method discovery and health checking via
synth-ai scan - First-Class Codex Support: Synth models now work seamlessly with Codex CLI workflows
- Multi-stage pipeline optimization for both MIPRO and GEPA algorithms
- GEPA Algorithm: Genetic Evolution for Prompt Optimization now available
- Google Gemini model support (
gemini-2.5-pro,gemini-2.5-flash,gemini-2.5-flash-lite) - Expanded OpenAI model support (GPT-4o, GPT-4.1, GPT-5 families)
- Prompt ID-based URLs for better traceability and concurrency
- Full terminal streaming logs for both SFT and RL training
- Hosted Synth judges with per-job overrides and rubric selection
- Rubric-aware filtering for SFT pipelines
- Qwen-VL support across SFT & RL with built-in vision collators
- Real-time status updates, detailed event logs, and full metrics logging
- Qwen Coder models supported across SFT and inference workflows
- SDK migrated to Turso for concurrency and higher throughput
- More training topologies on H200s with additional tensor/pipeline/data parallel layouts
- Full LoRA support for Policy Gradient and pipelined RL async rollouts
- Expanded Qwen catalog with SFT and inference presets for all Qwen releases
- Large-model inference & training topologies (2×, 4×, 8× layouts)
- LoRA-first SFT as first-class training mode across all Qwen topologies
- Turnkey rollout with automatic SKU surfacing
- Enhanced rollout viewer with real-time metrics and progress tracking
- B200 & H200 GPU support for training and inference workloads
- Faster inference with optimized pipeline and improved throughput
- GSPO (Group Sequence Policy Optimization) algorithm integration
- Organization-scoped environment credentials with sealed-box encryption
- First-party Task App integration with authenticated rollouts
- Single-node, multi-GPU Online RL with configurable tensor parallelism
- Production run flow with monitoring and checkpoint inference
- Fine-tuning (SFT) endpoints available and documented end-to-end
- Interactive demo launcher (
uvx synth-ai demo) with finetuning flow for Qwen 4B - Live polling output during training with real-time status updates
- New backend balance APIs and CLI for account visibility
- CLI utilities:
balance,traces, andmancommands
- Environment Registration API for custom environments
- Turso/sqld daemon support with local-first replicas
- Environment Service Daemon via
uvx synth-ai serve - Initial development release
- Cuvier Error Search (deprecated)
- Langsmith integration for Enterprise partners
- Python SDK v0.3 (simplified API, Anthropic support)