Descriptions:
Fahd Mirza puts Poolside’s newly released Laguna XS 2.1 through a live evaluation using the Hermit agentic framework. The model is a 33 billion total parameter mixture-of-experts architecture with just 3 billion active parameters per token, a 256K context window, and native interleaved reasoning between tool calls — purpose-built for long-horizon agentic coding work. Poolside trained it entirely in-house using their own model factory, an MoE optimizer (not the more common AdamW), FP8 KV cache, and async on-policy reinforcement learning. A companion open-weight DFlash speculative drafter model was released alongside it, reportedly doubling tokens-per-second on local inference. The model requires approximately 280GB of H100 or A100 VRAM to run locally.
The video covers three test tasks with candid results. First, debugging a broken full-stack call center dashboard app with scattered frontend and backend bugs — the model successfully identified and fixed all four issues. Second, generating a physics-based HTML animation of a drone view over an ocean with falling balls — the model failed, producing random rainbow circles with no ripple effects or ball physics. Third, translating concurrent Go code using goroutines and channels into nine supported programming languages — partially successful, with multilingual code generation being the headline benchmark improvement Poolside highlighted over its predecessor XS 2.2.
Published benchmark comparisons against Qwen 3.6 and Claude Haiku 4.5 on SWEBench Verified and TerminalBench are discussed, with XS 2.1 trailing on both despite its lower active parameter count. Released under the Open MDW 1.1 license, Laguna XS 2.1 is positioned as an open-weight alternative for developers running local agentic coding pipelines.
📺 Source: Fahd Mirza · Published July 02, 2026
🏷️ Format: Review
