Private, custom AI models trained on your enterprise data, benchmarked against your standards, and hosted via API. Managed with dashboards, monitoring and automation. Frontier-level intelligence for your workflows, at a fraction of the operating cost.
Train on your enterprise data, evaluate against your benchmarks, serve as private endpoints and own every checkpoint. Through the dashboard or the API.
Twenty-five open models across language, vision, image, video and audio.
Pick your data, choose a model, press start.
Models
Connect a source once and the training file keeps itself current.
Connect your sources12 connectors
Synced hourly 04:00
Nucleus fetches your data and generates training-ready datasets.
Test every checkpoint against the outcomes your operations require.
Benchmark scorethreshold 93.0
Business scenarios1,000 cases
Change one line. Your OpenAI or Anthropic client works as-is.
Base URLhttps://api.nucleus-ai.io/v1
Provisioned throughput1,240 of 2,000 rpm
from openai import OpenAI client = OpenAI( base_url="https://api.nucleus-ai.io/v1", api_key=NUCLEUS_API_KEY, ) response = client.chat.completions.create( model="ft:llama-3.3-70b:acme", messages=[ {"role": "user", "content": prompt} ], )
Take the weights the endpoint is serving and run them on hardware you control.
Also exports as
GGUFLoRA adapterMerged FP16Serve it anywhere
Agents that run on your fine-tuned models and your data.
Runbook Copilot
Tier 1 support
ft:llama-3.3-70b
Escalation Triage
Tier 2 support
ft:qwen3-32b
Pick a base model, connect your data, and deploy on a private endpoint. Dashboard or API, your first fine-tune runs in minutes.
Bring your JSONL or connect a data source. Pick a base model, and managed GPUs handle the rest.
Upload your test sets, build them in the dashboard, or let Nucleus generate them. Every checkpoint gets a domain score.
Swap your OpenAI or Anthropic base URL. The rest of your code stays the same.
Training data stays in your account. Weights are yours to export and run wherever policy requires.
Bring your JSONL or connect a data source. Pick a base model, and managed GPUs handle the rest.
Upload your test sets, build them in the dashboard, or let Nucleus generate them. Every checkpoint gets a domain score.
support-eval.jsonl4 criteria · pass ≥ 93.0
1,200 test casesReady
Score per checkpointBest · ckpt 3
Swap your OpenAI or Anthropic base URL. The rest of your code stays the same.
from openai import OpenAI client = OpenAI( base_url="https://api.openai.com/v1", api_key=NUCLEUS_API_KEY, )
{ "category": "billing", "priority": "high", "route": "tier-2", "action": "Refund duplicate charge" }
Training data stays in your account. Weights are yours to export and run wherever policy requires.
POST /v1/training-runs/run_3Bz /checkpoints/ckpt_000675/export {"format": "merged-safetensors"} 202 Accepted {"operation": "op_91c4"}
Also exports as
LoRA adapterGGUFDownloadOpen models across language, vision, image, video and audio. Fine-tune on your data, serve the result as a private endpoint, and keep the weights.
zai-org/GLM-5.3
MoE · 780B params · 40B active
Qwen/Qwen3-VL-235B-A22B-Instruct
MoE · 235B params · 22B active
black-forest-labs/FLUX.2-dev
Flow transformer · 32B params
Lightricks/LTX-2
DiT · native audio + video
Qwen/Qwen3-Embedding-8B
Bi-encoder · 8B params · 4096 dims
openai/gpt-oss-120b
MoE · 117B params · 5.1B active
Three meters run the bill: training tokens, served tokens and storage. Reserve dedicated GPUs when you need guaranteed capacity. Idle time costs nothing.
Custom models trained on Nucleus that connect to the tools a business already uses, and deliver the result a team used to spend a day on. Every workflow is production-ready and runs on your data.
47 trucks rerouted while they were still moving GPS traces, traffic and weather become live route changes, fuel efficiency per vehicle, and an SMS to every customer whose window slipped.
Five API calls take a JSONL file to a fine-tuned model. Two of them are the point: one serves it behind an OpenAI-compatible endpoint, the other hands you the weights. Prefer to drive the training loop yourself? Forward-backward passes and optimiser steps are first-class API calls too.
Bring a JSONL file of examples. That is the only prerequisite, and the file id it returns is what every later call refers to.
FilesFine-tune, evaluate, serve, and export models your organisation owns, with a workflow for every team that touches them.
Fine-tune and deploy custom models on your enterprise data with full control over training pipelines and inference endpoints.
Drive the training loop call by call. RL, DPO, distillation, and custom losses, with the GPUs handled for you.
Ship features on models that speak your domain. Keep your existing OpenAI or Anthropic client code and swap the base URL.
Turn proprietary datasets into evaluated, versioned models with managed fine-tuning jobs. Upload a JSONL file, get back a private endpoint.
Serve models behind OpenAI- and Anthropic-compatible endpoints, with usage tracking and rate limits built in.
Own your AI strategy. Training data stays in your account, weights are exportable, and models can run wherever policy requires.
Nucleus handles training and inference at any scale, from the first fine-tuning job to full production traffic, so your team ships models instead of managing GPUs.
100% yours
OpenAI + Anthropic
