Nebius Token Factory

Nebius Token Factory is an inference API for open-weight language models, with fine-tuning and dedicated endpoints, from Nasdaq-listed Nebius. Billed per token.

Nebius Token Factory gives developers hosted access to open-weight model families such as DeepSeek, Qwen, Kimi, Nemotron, GLM and gpt-oss through one API. Requests follow the OpenAI format, so an existing app changes its base URL and key to switch over. Nebius Group N.V., the Amsterdam AI cloud company listed on Nasdaq, builds and runs it. Token Factory keeps its own site, sign-up, documentation and billing, apart from Nebius AI Cloud. The Model catalog shows each model's context window, licence, processing region and per-token rate. Inference covers serverless endpoints, dedicated endpoints on isolated GPUs, a Playground, Prompt presets and Observability. Calls can use functions, follow a supplied JSON schema, read images and set a reasoning effort, and separate endpoints return embeddings and rerank scores. Post-training runs supervised fine-tuning with LoRA or full weights, plus custom speculator training. Data Lab keeps datasets and chat completion logs for reuse, and Sandboxes give software-engineering agents isolated environments to work in. Token Factory suits developers and machine learning engineers who want open models in production without operating GPU clusters themselves. Startups usually begin on serverless endpoints, paying per token at a rate set for each model, with input and output counted separately. Teams that need predictable latency, isolation or a fixed region move to dedicated endpoints, charged while their replicas run. Larger companies sign an Enterprise agreement through sales for higher rate limits and reserved GPU capacity. Typical jobs are chat and coding assistants, retrieval over company documents, and pulling structured fields out of text and screenshots. A team can keep the OpenAI SDK it already uses and wire retrieval through LangChain against the same endpoint. Each catalog model links to its weights on Hugging Face, so a model tried here stays portable to other hosts. Coding teams pair Sandboxes with Mini-SWE-Agent when an agent needs its own place to run code. Token Factory fits teams that choose a model per task and expect to fine-tune it or move it to dedicated capacity later.