Poolside has just released technical reports on Laguna M.1 and Laguna XS.2, two foundation models based on the Mixture-of-Experts (MoE) architecture designed specifically for agentic coding tasks.
Developments
The M.1 model features a total of 225.8 billion parameters (with 23.4 billion active parameters per token), while the more compact XS.2 version has 33.4 billion parameters (with 3 billion active parameters). Both were trained from scratch within poolside's tightly integrated "Model Factory" system.
In software and terminal benchmarks (SWE-bench Verified, Terminal-Bench 2.0), Laguna M.1 and XS.2 demonstrate strong competitiveness against leading open-source models in their respective weight classes. Notably, the weights of Laguna XS.2 have been publicly released under the Apache 2.0 license on Hugging Face.
Why It Matters
The launch of specialized MoE models for coding significantly reduces computational costs while maintaining high performance. For the Vietnamese developer community, the release of Laguna XS.2 (the open version) is a valuable option for self-hosting local code-assistant AI agents, rather than relying entirely on proprietary APIs.