FriendliAI Expands Model Deployment Options With Support for Poolside’s Laguna XS.2

According to a recent LinkedIn post from FriendliAI, the company is highlighting one-click deployment of Poolside’s new Laguna XS.2 open-weight agentic coding model on Friendli Dedicated Endpoints. The model is described as a 33B-parameter Mixture-of-Experts system targeted at agentic coding and long-horizon tasks, activating 3B parameters per token for inference efficiency.

The post notes four precision variants—BF16, FP8 with KV cache for Hopper-class GPUs, INT4 for minimal memory use, and NVFP4 optimized for NVIDIA Blackwell hardware—suggesting a focus on compatibility with current and next-generation GPU architectures. Performance metrics cited on SWE-bench benchmarks, along with features such as interleaved reasoning and a 256-expert design, indicate an emphasis on high-end developer and enterprise workloads.

From an investor perspective, this integration suggests FriendliAI is positioning its infrastructure as a flexible deployment layer for advanced third-party models, rather than relying solely on proprietary models. This could expand its addressable market among AI engineering teams seeking managed, performance-optimized endpoints for complex coding agents.

Support for multiple quantization formats and autoscaling dedicated GPUs may help FriendliAI target cost-sensitive and latency-sensitive use cases, potentially improving utilization of its infrastructure and supporting higher-margin, value-added services. If adoption of Laguna XS.2 or similar models grows, FriendliAI could enhance its competitive standing in the AI inference and model-serving segment, though the post does not provide information on pricing, customer traction, or revenue impact.