Laguna XS 2.1 AI Evaluation | LayerLens - Benchmark Results & Performance

Processing your request

This may take a moment. We're preparing the latest results.

Evaluations Compass Benchmarks Spaces

Learning Docs

Processing your request

This may take a moment. We're preparing the latest results.

Laguna XS 2.1 AI Evaluation | LayerLens - Benchmark Results & Performance

Model

Laguna XS 2.1 by Poolside

Text->text

Free credits included. No credit card required.

Run your own evaluation

Laguna XS 2.1 is Poolside's lightest and fastest agentic coding model, featuring a 33B total parameter Mixture-of-Experts architecture with only 3B activated parameters per token. It is purpose-built for agentic coding and long-horizon software engineering work, combining tool calling and native interleaved reasoning capabilities with a compact footprint suitable for local deployment. The model uses mixed sliding-window and global attention across 40 layers with sigmoid per-head gating and FP8 KV cache, supporting a 256K-token context window and up to 32K output tokens. Trained in-house on 15 trillion tokens, it is Poolside's first open-weight model release.

Evaluations

across 8 benchmarks

Latency

59.9s

Context length

262k

tokens

Cost

$0.06 · $0.12

input · output per 1M tokens

Key takeaways

The Laguna XS 2.1 model is designed for agentic coding and long-horizon software engineering tasks, featuring tool calling and native interleaved reasoning.
It is a 33B total parameter Mixture-of-Experts (MoE) architecture with 3B activated parameters per token, using mixed sliding-window and global attention across 40 layers, and was trained on 15 trillion tokens.
The model demonstrates strong mathematical problem-solving skills, particularly in basic and intermediate algebra, achieving 83% accuracy on the MATH benchmark subset. It supports a 256K-token context window and up to 32K output tokens.
It struggled with higher-level geometry and number theory on the MATH benchmark, and achieved only 26.67% accuracy on the AIME 2025 benchmark, indicating challenges with advanced geometric insight, precise combinatorial reasoning, and intricate algebraic execution.
Common failure modes include misinterpreting problem constraints, arithmetic errors in multi-step calculations, and incorrect application of mathematical properties, especially in complex geometry and number theory problems.

Technical specifications

Context length: 262k tokens
Max output: —
Architecture: MoE
Modalities: text->text
Parameters: ~33B
Open weights: Yes
Licence: Open
Regions: USA
Vendor: Poolside
Provider model key: poolside/laguna-xs-2.1