Baseten Launches Training Platform to Own Model Weights

Baseten, the San‑Francisco‑based AI infrastructure firm valued at $2.15 billion, has shifted its focus from inference to training with the launch of Baseten Training. The new platform promises to remove the ‘SSH‑in‑Friday, check‑on‑Monday’ pain by automating GPU cluster provisioning, multi‑node orchestration, and checkpointing on a multi‑cloud basis. By letting customers keep full ownership of their fine‑tuned weights, Baseten counters the lock‑in tactics of other training providers and positions itself as a full‑stack partner for enterprises moving away from OpenAI’s API dominance.

The core of Baseten Training is its Multi‑Cloud Management (MCM) system, which dynamically allocates NVIDIA H100 or B200 GPUs across 'AWS', 'Azure', 'GCP', and edge providers, allowing sub‑minute job scheduling without long‑term contracts. Integrated observability tooling offers per‑GPU metrics, granular checkpoint tracking, and a refreshed UI that surfaces infrastructure events in real time. Baseten also supplies an ‘ML Cookbook’ of open‑source recipes for models such as Gemma, GPT‑OSS, and Qwen, lowering the barrier to training success. Early adopters like Oxen AI and Parsed report 84 % cost savings and 50 % latency reductions, demonstrating the platform’s practical impact on custom model workloads.

Beyond the immediate cost and performance gains, Baseten’s strategy reflects a broader industry shift. As open‑source models improve and reinforcement learning or supervised fine‑tuning techniques mature, enterprises seek to replace expensive API calls with proprietary, domain‑specific models. By owning the training and inference lifecycle, Baseten can continuously iterate on models, even training draft models for speculative decoding that accelerate inference. The roadmap now points to deeper integration of advanced techniques, expansion into image, audio, and video fine‑tuning, and abstractions that capture common training patterns without sacrificing flexibility. In a crowded market, Baseten’s focus on developer experience, multi‑cloud agility, and weight ownership may well become the differentiator that keeps enterprises from falling back into closed‑source lock‑ins.

Key takeaway: Baseten’s emphasis on retaining model weight ownership and low‑level training infrastructure gives enterprises the flexibility to fine‑tune open‑source models while still leveraging the company’s inference expertise, creating a more sustainable path away from proprietary AI services.