Managed post-training API from Thinking Machines Lab (Mira Murati's startup). Targets researchers and developers who want full control over fine-tuning loops, SFT, GRPO/PPO RL, and DPO without managing GPU clusters. Uses LoRA to batch multiple training runs across shared compute, covering models from Qwen3-4B to large MoEs like Kimi-K2.
Thinking Machines rolled out Dimuon, a LoRA-specific optimizer that beats AdamW on step count in supervised fine-tuning while leaving RL mostly unchanged.
Bridgewater and Thinking Machines show that expert-labeled data plus on-policy distillation beats every frontier model on financial document triage — at 13.8x lower cost.