Model Spend Arena2ND ED.
33B · MoE · 262,144 ctx · 2026-08-14

Laguna XS.2

Runs on ≤32 GB · Hugging Face ↗

vLLMSGLang

Not independently scored by Artificial Analysis

Poolside's first open-weight coder — 33B total / ~3B active MoE, Apache-2.0, built for agentic long-horizon coding. Reportedly 68.2% SWE-bench Verified and beating dense models 10x its size; the local-runnable member of the Laguna family (S 2.1 / M.1 are data-centre). Custom `LagunaForCausalLM` arch.

First-party test · not the AA coding index

BigCodeBench-Hard pass@1 34% (41/121), via rented RTX PRO 6000 Blackwell 96 GB (vLLM, BF16) — comparable full-148, conc64. The brutal counterpart to HumanEval — where HumanEval saturates near the top, BCB-Hard spreads the field, so this is the number that actually separates coding ability. The local-runnable Laguna — lands with the strong open coders on hard problems; agentic by design, so a single-shot benchmark is a floor for its long-horizon SWE behaviour.

Sizes on disk

Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈5 GB at 32K, fp16).

QuantizationSize
Q3_K_M15.6 GB
IQ4_XS18.2 GB
Q4_019.2 GB
Q4_K_S19.8 GB
Q4_K_M20.5 GB
Q420.7 GB
Q4_121.2 GB
Q5_K_M24.0 GB
Q2_K24.5 GB
Q8_035.6 GB
Q6_K58.2 GB
BF1666.9 GB

Config tips

~20 GB at Q4 — a 24 GB+ card or the 32 GB tier. Also on OpenRouter (laguna-xs-2.1:free) and opencode Zen's free set, so you can try the family for $0 before committing local VRAM.