Laguna XS 2.1 on 24GB vs 32GB VRAM (July 2026): The Context Ceiling Is the Real Difference

Bottom Line (July 2026)

• What it is: Laguna XS 2.1, Poolside, released July 2, 2026. 33B total params, 3B active, Mixture-of-Experts, 256K native context, ~70.9% on SWE-Bench Verified.
• It fits both tiers. laguna-xs-2.1:q4_K_M in the Ollama library is 20.27GB — loads on a 24GB 3090/4090 and on a 32GB 5090.
• The real difference is context. 24GB leaves ~4GB for KV cache (8–16K practical context). 32GB leaves ~12GB (~64K).
• Q8 fits neither card. The Q8 weights are ~33–36GB before KV cache. A 32GB card does not unlock Q8.
• Speed is small-model speed. Only 3B params active per token: 45–60 tok/sec on a 5090.
• 64GB+? Step up to Laguna S 2.1 (118B/8B, 1M context) instead.

It Fits Both Cards — So What Are You Actually Buying?

After Poolside shipped Laguna XS 2.1 on July 2, the question that shows up in every thread is some version of “does it fit my card?” The answer is yes for both 24GB and 32GB — which makes it the wrong question.

The right question is what each tier leaves you for KV cache, because the KV cache is where your context window lives. Weights are fixed; context is what scales with the leftover VRAM.

The model’s native window is 256K. Neither consumer card reaches it — the card, not the model, is the ceiling at both tiers.

Two Things People Get Wrong

1. Q8 does not fit 32GB. The Q8 weights are roughly 33–36GB, before you count a single token of KV cache. We keep seeing the assumption that a 32GB card unlocks Q8 because the VRAM number went up. It does not. On a 5090, stay on Q4_K_M and spend the extra ~12GB on context instead — that trade is heavily in your favor for coding work.

2. Speed does not scale with total params. Laguna XS 2.1 is 33B total but only 3B active per token, so generation runs at small-model speed — around 45–60 tok/sec on a 5090. Do not skip it because “33B sounds slow on my card.”

Why the Context Gap Matters for Agentic Coding

If you run agentic coding loops — OpenClaw, Claude Code with a local backend, aider — the 24GB → 32GB step buys you context, not model size.

An 8–16K window forces the agent to constantly re-feed the repo: every few turns it drops files it already read and pulls them back in. That burns tokens, burns time, and degrades the agent’s grasp of the codebase mid-task. A ~64K window holds a real working set — several source files, the diff, and the conversation — without thrashing.

That is a bigger practical difference than the tok/sec gap between the cards, and it is the honest reason to pick a 5090 for this model.

On 24GB and want long context anyway? Your options are: cap the window and accept the re-feeding, or step to 48GB+ — at which point Laguna S 2.1 (118B/8B, 1M context) enters the picture at UD-Q2_K_XL.

Quick Setup

ollama pull laguna-xs-2.1:q4_K_M

On 24GB, cap the context explicitly so the KV cache does not spill to system RAM:

OLLAMA_CONTEXT_LENGTH=16384 ollama serve

On a 5090, 64K is the sweet spot:

OLLAMA_CONTEXT_LENGTH=65536 ollama serve

If generation suddenly crawls, the cache spilled — drop the window one notch. More on that failure mode in why local LLMs are slow even when they fit.

Related Guides

• Best local LLM for the RTX 5090 — full 32GB tier picks
• Best local LLM for the RTX 3090 and RTX 4090 — 24GB tier picks
• Laguna S 2.1 local setup — the 118B big sibling for 64GB/128GB
• RTX 5090 vs 4090 vs used 3090 — which card to buy