​
July 2025

​
Serverless Fine-tuning

Serverless Fine-tuning
Introducing Serverless Fine-tuning – the fastest way to customise open-weight foundation models without touching infrastructure. Spin up a secure, pay-as-you-go training job in a single API call, watch metrics stream in real-time, and download your tuned model or push straight to Hugging Face.

​
Features

  • Two-step workflow. Pick any supported base model (Llama 3, Mistral 7B, DeepSeek, Qwen and more) and launch a job with your dataset – no cluster sizing, no Dockerfiles
  • LoRA-powered efficiency. Default Low-Rank-Adaptation (LoRA) reduces GPU hours and cost
  • Live metrics & easy monitoring. Poll one endpoint to track train_loss, eval_loss, perplexity
  • Export or deploy instantly. One-click push to Hugging Face or direct download of a ready-to-serve artefact
  • Serverless pricing. $2 minimum per job, billed by processed tokens; every new account still gets $5 free credit to experiment

​
Quick start

# 1. List available base models
curl -H "Authorization: Bearer $NSCALE_API_TOKEN" \
     https://fine-tuning.api.nscale.com/api/v1/organizations/$ORG_ID/base-models

# 2. Launch a LoRA job
curl -X POST https://fine-tuning.api.nscale.com/api/v1/organizations/$ORG_ID/jobs \
     -H "Authorization: Bearer $NSCALE_API_TOKEN" \
     -H "Content-Type: application/json" \
     -d '{
           "name": "support-bot-finetune",
           "base_model_id": "e5f6a7b8-c9d0-1234-efab-567890123456",
           "dataset": {"id":"<DATASET_ID>","prompt_column":"prompt","answer_column":"response"},
           "hyperparameters":{"n_epochs":3,"batch_size":4,"lora":{"enabled":true,"r":8,"alpha":16}}
         }'

# 3. Stream training metrics
curl -H "Authorization: Bearer $NSCALE_API_TOKEN" \
     https://fine-tuning.api.nscale.com/api/v1/organizations/$ORG_ID/jobs/$JOB_ID/metrics
​
March 2025

​
Serverless Inference

Serverless Inference
Our fully-managed, pay-per-request runtime that puts a pool of GPUs behind a single OpenAI-compatible endpoint. Instead of capacity planning, container images and infra dashboards, you call https://inference.api.nscale.com/v1/* and get deterministic, low-latency responses from today’s best open-source models—all billed per token and delivered from data-sovereign, 100% renewable data-centres.

​
Features

  • OpenAI-compatible endpoints. Drop-in support for Llama, Qwen, DeepSeek and other leading models makes migration a copy-paste job
  • Pay-as-you-go billing. Prices are per 1 million tokens including input and output tokens for Chat, Multimodal, Language and Code models. Image models is based on image size and steps
  • 80% lower cost & 100% renewable. Our vertically-integrated stack slashes TCO versus hyperscalers while guaranteeing data privacy—requests are never logged or reused
  • $5 free credits to get started. Every new account includes starter credits so you can ship to production in minutes

​
Under the hood

AreaWhat it looks likeWhy it matters
API surfaceDrop‑in equivalents for GET /models, POST /chat/completions, POST /images with optional stream: true for SSE (text/event-stream).Migrate from OpenAI by changing only the base URL and key.
Model libraryLaunch set covers Meta Llama‑4 Scout 17B, Qwen‑3 235B, Mixtral‑8×22B, DeepSeek‑R1 distills, SD‑XL 1.0 and more (text, code, vision).Lets teams A/B models or mix modalities without provisioning extra infra.
Elastic runtime“Zero rate limits, no cold starts.” Traffic is sharded over thousands of MI300X/MI250X/H100 GPUs, spun up on‑demand by our orchestration layer.Bursty workloads stay < 200 ms tail latency without you over‑allocating GPUs.
Cost modelTokens in, tokens out — billed per 1M tokens; images billed per megapixel. Every account starts with $5 free credit.Fine‑grained, deterministic spend; easy to embed in metered SaaS.
Security / privacyEnd‑to‑end TLS, org‑scoped API keys, full tenant isolation; we never log or train on user prompts or outputs.Meets GDPR, HIPAA and most vendor‑assessment checklists out of the box.
SustainabilityAll compute runs in hydro‑powered facilities; the vertical stack is 80% cheaper‑per‑token than hyperscalers.Fewer carbon (and budget) emissions per request.

​
Quick start

curl -X POST \
  https://inference.api.nscale.com/v1/chat/completions \
  -H "Authorization: Bearer $NSCALE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
        "model": "meta-llama/Llama-4-Scout-17B-Instruct",
        "messages": [{"role":"user","content":"Hello world"}],
        "stream": true
      }'