Tinker vs Unsloth

Tinker

Full control of the training loop with the GPUs abstracted away, plus recipes for SFT, DPO, RL and distillation. LoRA only; no full-parameter training.

Unsloth

The Apache-2.0 core runs on customer hardware and keeps model weights there. Users supply and pay for the GPU.

Tinker

1. Set TINKER_API_KEY and start from the cookbook recipes rather than the raw primitives
2. Read the 'Avoid Client-Side Timeouts and Retries' guide before wrapping sampling calls in your own retries; the SDK already retries sampling with stable request IDs
3. Save intermediate checkpoints with a TTL between 1 hour and 10 years; storage bills at $0.10 a GB-month until they expire
4. Read models.json for current prices before a run; sampling tokens cost more than training tokens on the open models
5. Check the model deprecations page before pinning a base model; 18 were retired on 2026-06-12

Unsloth

1. Install with uv pip install unsloth --torch-backend=auto on a CUDA machine; the desktop app is for people
2. Start from the notebook for the model family in unslothai/notebooks; it sets LoRA targets and the chat template
3. Save the LoRA adapter while iterating and merge to 16-bit or GGUF only when you ship
4. If Studio must be reachable by other agents, pass --disable-tools and keep it on 127.0.0.1 behind a tunnel
5. Pin the exact unsloth version; releases land several times a week and don't flag breaking changes