Poolside's Laguna S 2.1: A Small, Cheap, Open Coding Model

Poolside has released its new coding model, Laguna S 2.1, as open source. It's small — 118 billion parameters, only 8 billion active — and Poolside calls it "the most capable agentic coding model in its weight class." In other words, it competes with similar-sized models, not the giants. The interesting part is the company's thesis: their goal wasn't to add "more intelligence" but to make the model more careful — to verify, not declare victory early, and be persistent.

In brief

• What launched: Laguna S 2.1, Poolside's coding-focused open model. 118B/8B-active (MoE), 1M-token context. Weights open on Hugging Face.
• Where it stands: good for its size class but behind the frontier — e.g. 70.2% on Terminal-Bench 2.1 vs Kimi K3's 88.3% and Fable 5's 88.0%. Poolside publishes those gaps openly on its own page.
• The draw — price and access: on OpenRouter, $0.10 input / $0.20 output per 1M tokens (very cheap); free to try with no login at chat.poolside.ai; and it runs locally via Ollama and vLLM.
• What's different — candor: Poolside makes all its evaluation trajectories public (trajectories.poolside.ai) and lists the model's shortcomings itself (overthinking, JSON-escaping issues).

Our take

If you code or build something, this is a genuinely usable option for you: try it free with no login, run it on the API for what's practically pennies, or run it on your own machine. But set expectations — it doesn't beat the frontier models and doesn't try to; it's a cheap tool that does a lot for its small size. The real value is beyond price: Poolside says it made the model "not smarter, but more careful" — a model that doesn't call "done" too early and checks its work is often more useful in coding than raw intelligence — and it publishes all its test records, including its flaws. When picking a model, that second part is far more reassuring than a shiny score.

Kaynak: Poolside