Python SDK for LoRA fine-tuning with supervised fine-tuning and reinforcement learning
Compress prompts and reduce LLM API costs while preserving answers
Stage
Not given
Startup
Founded
Not given
2024
Based in
Not given
Not given
Pricing model
Not given
Free
Pricing
Not given
Open beta: $0/month with unlimited API calls, all features, and no credit card required. After beta: a generous free tier and simple usage plans with early users keeping their perks.
Key features
forward_backward() operation for computing loss and accumulating gradients
optim_step() for updating the adapter
sample() for generating from current or saved policy
save_state() to save and resume training state
Tinker API compatibility
Separate trainer and sampler resources
Synchronous and asynchronous RL patterns
Frozen base model weights with LoRA adapter updates
Support for open models from Qwen, Google, DeepSeek, Moonshot AI, NVIDIA, and Z.ai
Compress long prompts, logs and documents by removing filler while preserving code, numbers and facts
Convert web pages to clean Markdown instead of raw HTML
Shrink JSON responses into TOON format with headers and CSV-like rows
Available as MCP server for Claude, Cursor, Windsurf and VS Code
Gateway integration for OpenAI and Anthropic SDKs
REST API for direct integration
Self-hosted or hosted gateway options
Dashboard showing token and cost savings
Works with any chat model including local models
Output modes to control answer verbosity: lite, caveman and ultra
What makes it different
Managed training and sampling infrastructure, no setup required
Tinker-compatible API for easy migration
Ready-made SFT and RL recipes through Loops cookbook
Support for vision-language models
Rule-based compression takes milliseconds without using another LLM
Your API key never leaves your infrastructure when using self-hosted gateway
Streaming, tool calling and function calling work unchanged
Real-world savings up to 42% on monthly LLM bills according to calculator
Loops
AgentReady