Skip to content

Laguna XS 2.1

OpenMDW-1.1

Poolside · 33B · transformer-moe

2026-07-02262K context33B params

Use Cases

chatcodereasoningtools

Quantization Options

QuantBitsVRAMQualityStatus
Q4_K_Mrec422.0 GBGood
Q8_0839.0 GBExcellent
BF161672.0 GBExcellent

About this model

Laguna XS 2.1 is Poolside's small agentic coding model, released July 2026 under the permissive OpenMDW-1.1 license and explicitly designed for long-horizon coding work on a local machine. A 33B-total / 3B-active Mixture-of-Experts, it scores 70.9 on SWE-bench Verified — remarkable for its size — with native tool use and interleaved thinking between tool calls. The architecture is tuned for local efficiency: 30 of its 40 layers use sliding-window attention and the KV cache is FP8-quantized, keeping memory low across the 256K context window. The 20 GB Q4_K_M (the default Ollama tag) fits a 24 GB GPU or a 36 GB Mac, and with only 3B active parameters per token it generates fast even with partial CPU offload.

Benchmarks

70.9
swe-bench-verified
47.6
swe-bench-pro