Fine-tuning AI apps

Managed fine-tuning API with low-level control over the training loop.

Tinker is Thinking Machines Lab's training API for fine-tuning open-weight LLMs. It exposes low-level primitives — forward_backward, optim_step, sample — so researchers keep full control of data and algorithms while the service handles distributed GPU scheduling and failure recovery. LoRA-based runs cover models from small Llamas up to large mixture-of-experts like Qwen-235B and Kimi K2, and trained weights can be downloaded.

Open-source post-training for LLMs — LoRA to RL, all from one YAML config.

An open-source (Apache-2.0) framework that streamlines post-training for open-weight models: full fine-tuning, LoRA/QLoRA, preference tuning (DPO, IPO, KTO, ORPO), reinforcement learning (GRPO), reward modeling and quantization-aware training, configured through a single YAML file with no scripting. Wraps Hugging Face Transformers, PEFT, TRL and DeepSpeed, and supports dozens of model families including multimodal vision and audio models.

Fine-tune open-source LLMs and serve them in production.

Predibase is an enterprise platform for fine-tuning open-source models and serving them in production. It pairs a post-training stack — supervised fine-tuning plus an end-to-end reinforcement fine-tuning (RFT) flow — with an optimized inference engine, and its open-source LoRAX framework serves many fine-tuned LoRA adapters from a single GPU. Runs as managed SaaS or inside your own VPC.

Fine-tune open LLMs 2x faster with far less VRAM. Open source.

An open-source (Apache-2.0) framework for fine-tuning and running open-weight models with custom CUDA kernels — roughly 2x faster training and large VRAM savings, so 7B–13B models fit on a single consumer GPU. Free tier runs on Colab/Kaggle or locally; Pro and Enterprise tiers add multi-GPU and multi-node speedups. Exports to GGUF/Safetensors for llama.cpp, vLLM, and Ollama.

Replace frontier-model spend with a fine-tuned small model.

Captures your production OpenAI / Anthropic calls, builds a dataset, fine-tunes a small open-weights model on your traffic, then serves the swap behind your existing SDK. The pitch: 10x cost reduction at parity.