Summary
Fahd Mirza puts Laguna S2.1 — a newly released 118-billion-parameter mixture-of-experts model from Poolside — through its paces in a live agentic coding evaluation. The model features 8 billion active parameters per token, a 1-million-token native context window, and is purpose-built for long-horizon agentic tasks. Mirza runs it through Hermes agent using API access and gives it a deliberately under-specified debugging task: find and fix whatever is wrong in a full-stack C-port freight tracker application, without being told what the bug is.
The model successfully identifies and fixes the actual issue — a vessel draft validation bug that allowed ships to be assigned to berths with insufficient depth clearance — but the test also reveals significant weaknesses. Laguna S2.1 hallucinates two additional bugs that do not exist, gets caught in self-referential reasoning loops, and explicitly acknowledges it is overthinking while still failing to exit the loop. Mirza flags token cost and latency implications of this behavior as meaningful concerns for production agentic use. For context, he references TerminalBench scores placing Laguna at 70.2, ahead of DeepSeek V4 Pro Max at 64 but well behind Claude Fable 5 at 88.
A second test asks the model to generate a photorealistic Greek souvlaki visualization as a single self-contained HTML file using pure Canvas and vanilla JavaScript — no WebGL, no libraries — with dynamic fire lighting, specular highlights, smoke, and 60fps rotation. The test probes the boundary of what single-file canvas rendering can realistically deliver, providing a useful signal for developers evaluating the model for creative code-generation workloads.
📺 Source: Fahd Mirza · Published July 22, 2026
🏷️ Format: Review
