FriendliAI
Inference cloud for open-weight models, from the team behind the ORCA paper
Overview
FriendliAI is an inference platform spun out of Seoul National University and now run from San Francisco, built on its founders' ORCA research into continuous batching. It sells per-token Model APIs for open models such as GLM-5.3 and DeepSeek-V3.2, Dedicated Endpoints billed per second on A100 to B300 GPUs, and BYOG deployment on a customer's own hardware.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- GLM-5.3 $1.40 input / $4.40 output per 1M
- DeepSeek-V3.2 $0.50 / $1.50 per 1M
- whisper-large-v3 at $0.0015 per audio minute
- H100 $5, H200 $7, B200 $9, B300 $12 an hour
- No charge for start-up time
- Pricing tailored to your infrastructure
- Custom rate limits and reserved GPU capacity
- Custom region, VPC and on-prem deployment
- Dedicated support and named customer success
Use cases
What you can produce with FriendliAI
- Open-weight models such as GLM-5.3, DeepSeek-V3.2 and Gemma 4 called through an OpenAI-compatible API
- Hugging Face or fine-tuned models deployed on dedicated A100 to B300 GPUs billed per second
- Inference run on your own GPUs with the Friendli stack under a BYOG contract
- Speech transcribed with whisper-large-v3 at $0.0015 per audio minute
- Coding agents connected to FriendliAI models with the FriendliLink CLI
- Rate limits that rise automatically from 100 to 10,000 requests a minute with lifetime spend
ASEAN Perspective
FriendliAI in Southeast Asia
FriendliAI has meaningful APAC roots and traction: founded as a Seoul National University spinout, it maintains a Seoul operational hub and counts South Korean majors SK Telecom, Scatter Lab (active across Korea and Japan), and Upstage as named customers. The exclusive API launch partnership for LG AI Research's EXAONE 4.0 model is a strong signal of its position as preferred inference partner for Korean AI labs. The $20M seed extension (August 2025) included Korean institutional investors KDB and KB Securities alongside US VCs, and the company has explicitly flagged Asia expansion alongside North America as a use of proceeds. ASEAN-headquartered teams will find a vendor with demonstrated Korean-market depth, OpenAI-compatible APIs, and a cost-reduction pitch that translates well to GPU-cost-sensitive emerging market AI deployments.
FriendliAI has real research pedigree and named production users: SK Telecom reports five times the throughput, and Scatter Lab runs a billion monthly chatbot interactions on it. Prices are published and easy to model, from $0.14 per million input tokens and $4 an hour for a dedicated A100, with an OpenAI-compatible API and SOC 2 Type II and HIPAA claims.
Treat the "up to 90% lower cost" line as marketing until you benchmark your own workload. It is a small, seed-funded vendor with prepaid billing and no free tier, so teams needing multi-region guarantees should confirm capacity with sales.
What we found
FriendliAI is an inference cloud for open-weight and custom models, spun out of Seoul National University in 2021 by Byung-Gon Chun. Its roots are the ORCA paper presented at OSDI 2022, which described iteration-level scheduling, the technique now generally called continuous batching. The company is now run as FriendliAI Corp from San Francisco, where it opened an office in May 2026, with its engineering hub still in Seoul.
There are three ways to buy. Model APIs charge per token for hosted models, from $0.14 per million input tokens for Gemma 4 31B to $1.40 input and $4.40 output for GLM-5.3. Dedicated Endpoints rent a GPU billed per second: since 1 October 2026 an A100 costs $4 an hour, an H100 $5 and a B300 $12, after promotional rates ended. BYOG runs FriendliAI's stack on a customer's own GPUs on contract terms. Billing is prepaid from $10, requests stop when credits run out, and rate limits rise with lifetime spend, from 100 to 10,000 requests per minute.
The marketing claims up to 90% lower inference costs and up to seven times faster output, with a 99.99% uptime SLA. Those are vendor figures. The case studies are more concrete: SK Telecom reports five times the throughput at a third of the cost, and Scatter Lab serves its Zeta chatbot at a billion interactions a month. FriendliAI says any of 620,000 Hugging Face models can be deployed, and it was the first public API for LG's EXAONE 4.0 in 2025, though its K-EXAONE endpoint was withdrawn in September 2026.
It is still small, funded by a $20 million seed extension in 2025, and growth figures come from the chief executive. In September 2026 it signed Singapore-founded Aolani for GPU capacity, a useful sign for Asian latency, though no data-centre locations were named. Teams leaving closed-model APIs get a straightforward OpenAI-compatible endpoint and a FriendliLink tool that connects coding agents such as Claude Code or Cursor.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- The ORCA paper published by FriendliAI's founding team at OSDI 2022 introduced continuous batching — the scheduling technique that now powers vLLM and most major open-source inference engines, meaning the entire industry runs on an idea from this team.
- Scatter Lab used Friendli Container to scale its Zeta conversational AI to over 1 billion monthly interactions across Korea, Japan, and the US — reaching financial breakeven in the process.
- FriendliAI supports over 430,000 models from Hugging Face and acted as the exclusive API launch provider for LG AI Research's 32-billion-parameter EXAONE 4.0 model.
- CEO Byung-Gon Chun holds a PhD from UC Berkeley and an MS from Stanford, yet returned to Seoul National University as a professor before spinning out FriendliAI — a rare case of a top Korean university producing a Silicon Valley-headquartered AI infrastructure company.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including FriendliAI's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with FriendliAI unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to FriendliAI directly →
Spotted something out of date? Suggest an update →
Alternatives to FriendliAI
More in Code & Dev Tools
