FriendliAI

FriendliAI is an inference platform that focuses singularly on running open-weight and custom AI models faster and cheaper than the competition. The team's research roots are in serving system performance — they're known for the original Orca paper on continuous batching, which became foundational technology across the industry — and the product capitalizes on that with custom GPU kernels, smart caching, speculative decoding, parallel inference, and other low-level optimizations that compound into 2x+ throughput at lower latency on the same hardware. The platform offers serverless endpoints for popular open models, dedicated endpoints for custom or fine-tuned models with predictable performance, and a container deployment option for customers who need to bring inference into their own VPC or on-prem. FriendliAI advertises 99.99% uptime SLAs backed by geo-distributed infrastructure and multi-cloud failover, which is a meaningful differentiator for production workloads where most cheaper inference providers have spotty availability. Customers tend to be growth-stage AI companies running large open-weight workloads where the cost-per-token math matters. Pricing follows the standard usage-based pattern for serverless, plus dedicated capacity pricing for predictable rate-limited workloads; enterprise plans add SOC 2, BYOC, and committed volume discounts.