Novita AI

What it is for

Novita is a model API + GPU cloud: call a hosted DeepSeek / Llama / Qwen endpoint, or deploy a dedicated replica, or rent a GPU instance. Use it when the job is one vendor for tokens and boxes, not when you need a visual design studio.

It is not open source as a product. Serverless DeepSeek V4.1 Flash is $0.30 / $1.20 per million input / output tokens. Dedicated H100 SXM is $1.99/hour.

What makes it unique: the cover is Novita’s official model catalog chrome — Llama 3.3 70B, Qwen2.5 72B, Gemma, Phi-4 tiles plus a Deploy in 3 steps strip (pick model → choose GPU → deploy).

Key features

• Serverless model APIs — OpenAI-shaped chat completions across 200+ listed models (Quickstart; Pricing).
• Dedicated endpoints — pin a model (or a custom / LoRA) to GPUs that scale to zero (Dedicated endpoints).
• GPU instances — rent a VM from the marketplace (4090 / H100 / …), SSH or expose a port (Create instances).
• Agent sandboxes — a third product tab on the pricing page for isolated agent runtimes (Pricing).
• Batch inference — submit an async batch against the same model catalog instead of a live chat loop (Pricing).
• Image / video / speech APIs — the same OpenAI-shaped host also lists media models next to the chat catalog (Quickstart).

Who it's for

Best fit: builders who will call one OpenAI-shaped URL and maybe later pin a dedicated H100.

Poor fit: teams that want a single-operator reserved 8× H100 cluster (Lambda).

Pricing

Verified 2026-10-22 (Asia/Shanghai) from novita.ai/pricing and Dedicated endpoints. Token rows move. Copy the model you will call.

Serverless examples (per million tokens)

• DeepSeek V4.1 Flash / Flash 1017 — input $0.30 / Mt, cache-read $0.006 / Mt, output $1.20 / Mt, 1M context.
• DeepSeek V4 Pro 0813 — input $1.32, cache $0.044, output $3.96.
• Llama 3.3 70B Instruct — $0.135 / $0.40.
• Llama 3.1 8B Instruct — $0.02 / $0.05.

Dedicated GPUs (per GPU-hour, per-second on running replicas)

• RTX 4090 — $0.61/hr
• H100 SXM — $1.99/hr
• H200 SXM — $2.99/hr

Scale-to-zero: idle dedicated endpoints do not bill. GPU instance disks: container excess $0.005/GB/day; network volume $0.002/GB/day. HTTP 200 and 499 (client disconnect after inference starts) are charged; 4xx-before-model and 5xx are not (LLM billing).

Limitations / costs

• No public $0 hosted token card. You need credit on the account.
• 499 still bills. Do not cancel a long generation to “save” tokens.
• Dedicated $1.99 H100 is not a marketplace min — instance Explore rows can differ.
• Not a Zap / design tool.

Alternatives

• Fireworks AI — hosted inference if you want another public token card.
• Together AI — open-weight APIs if you will not rent Novita GPUs.
• fal — one HTTP price row if the job is media, not DeepSeek chat.

Official tutorial

Official Novita AI channel exists. The current oEmbed we resolved (How Many GPUs May You Need, 6T4DORCF5_U) is a capacity explainer, not a console getting-started, so it is not embedded. Use the written Quickstart plus Pricing.

Website

https://novita.ai

Source

Catalog row curated for AI Tools Nav. Cover is Novita’s official model-catalog / Deploy-in-3-steps chrome. DeepSeek V4.1 Flash $0.30 / $1.20 per 1M. Dedicated H100 $1.99/hr. Official channel: @novita_ai (no getting-started embed).

Last verified

2026-10-22 (Asia/Shanghai) editorial check.