FriendliAI

What it is for

FriendliAI is an inference cloud: call a serverless Model API, or pin a Hugging Face (or uploaded) checkpoint to its own GPUs, or run the same engine on GPUs you already own. Use it when the job is serve tokens (or audio seconds) on published $/hr or per-token rates, not when you need a $20 chat seat.

What makes it unique: the cover is Friendli’s official Suite still — sidebar (Models, Serverless Endpoints, Dedicated Endpoints, Container, Uploaded Models) plus Create an endpoint — not a pricing card.

Key features

• Model APIs (Serverless Endpoints) — OpenAI-shaped HTTP on a curated model list; text is billed in processed tokens, audio in seconds (Model APIs; pricing). One serverless call; GLM / DeepSeek / Whisper names are SKU.
• Dedicated Endpoints — pin a Hugging Face repo (630k+) or your uploaded weights to reserved GPUs, then call api.friendli.ai/dedicated/v1 (Dedicated QuickStart; Models; Full Tutorial). A reserved box, not a shared token meter. LoRA adapters, Host KV cache, and the Playground tab sit on this endpoint — they are not extra products.
• Container / BYOG — run the Friendli engine on GPUs you already have (pricing BYOG). A stack you host, not a Friendli GPU hour.
• Usage and costs — Suite Settings Usage plus GET /v1/team/usage and GET /v1/team/cost for tokens, GPU-seconds, and rated dollars by model / product / user (Usage and Costs). A chargeback desk, not a generate.
• Datasets — Suite Labs → Datasets: create a dataset, add train/validation/test splits, upload conversational JSONL / Parquet samples (POST /beta/dataset) (Create a dataset; Create a split). A data store, not an endpoint. Fine-tune jobs that later consume a dataset sit on Dedicated.

The Python/OpenAI client and API_KEY are install / SDK. The Playground is ops / try-it. GPU-seconds (Dedicated) and tokens / audio-seconds (Model APIs) are the meter. GPU type names (A100 / H100 / H200 / B200 / B300) and Reserved-GPU commits are SKU of Dedicated.

Who it's for

Best fit: inference teams who will start on Model APIs (pay per token) and only open Dedicated when they need a pinned H100 at $5/hr — not a team looking for a $20 chat plan.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from friendli.ai/pricing and Models and Pricing. Dedicated prints $/hour (billed per second). Model APIs are per-model; the docs table is JS — do not invent a token $ we did not copy.

• Dedicated on-demand — A100 80GB $4.0/hr; H100 80GB $5.0/hr; H200 141GB $7.0/hr; B200 180GB $9.0/hr; B300 288GB $12.0/hr. No extra start-up charge. Rates updated 2026-10-01 (changelog).
• Model APIs — usage-based. Text: processed tokens. Audio: per-second, shown as a per-minute rate. Confirm the live per-model row on the pricing page / docs table before you buy.
• BYOG — custom (contact@friendli.ai).
• Enterprise — custom. Rate limits, reserved GPUs, VPC / on-prem, named CS.

Do not invent a $20 Friendli seat. Confirm checkout if the Dedicated GPU table has moved.

Limitations / costs

• Model API token dollars are per-model and JS-rendered. Copy the row at checkout; we only print Dedicated GPU hours here.
• Dedicated bills while the endpoint is up. Sleep or terminate it.
• A serve cloud, not a trainer. For YAML post-training, see Axolotl.

Alternatives

• Fireworks AI — another serverless + GPU inference desk when Friendli’s Dedicated table is not the job.
• Together AI — hosted inference when you already live on their catalog.
• Runpod — a GPU marketplace when you want raw hours, not Friendli’s engine.

Official tutorial

Official FriendliAI channel video: Friendli Dedicated Endpoints - Full Tutorial (oEmbed author: FriendliAI).

Written path: Dedicated QuickStart plus Pricing. Channel also has Model APIs – Quickstart Tutorial With GLM-5.2 (same official author).

Website

https://friendli.ai

Source

Catalog row curated for AI Tools Nav. Cover is Friendli’s official Suite still (sidebar: Serverless / Dedicated / Container + Create an endpoint). Dedicated GPU hours checked on friendli.ai/pricing (A100 $4; H100 $5; H200 $7; B200 $9; B300 $12 / hr). Tutorial oEmbed author: FriendliAI.

Last verified

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