← Leaderboard
poolside/laguna-xs-2.1

Laguna XS 2.1

vLLM MoE · 3B active / 33B total AgentsCode
Throughput @ 4k
21.4t/s
@ 50k fill
8.8t/s
@ 100k fill
6.3t/s
Retention 100k÷4k
29%
Engine
vLLM
Parameters
3B active / 33B total
Released
2026-06-20

Headline numbers are perfbench-metrics (PBM): decode tok/s after a fixed context fill (4k / 50k / 100k), same agent-style workload as bench v2. Retention is how much of the 4k speed remains at 100k.

Fill curve

Throughput vs context fill — how the model holds up as the window grows.

24 12 0 21.4 4k 8.8 50k 6.3 100k tok/s

Context ladder

Older bench-v2 / golden cells at each benched context window.

Context KV Throughput
@ 32k peak auto 13.7t/s
@ 128k auto 11.6t/s
@ 256k auto 11.7t/s

Golden profile

poolside-laguna-xs-2-1-dflash-eugr

Capabilities

codingagenticvllmeugrdflashmoe

Why we run it

Poolside Laguna XS 2.1 + DFlash sidecar — bench v2 golden 262k @ 11.7 tok/s 2026-07-08.

Bench notes

PBM 4k @ 21.4 tok/s — perfbench-metrics — profile=poolside-laguna-xs-2-1-dflash-eugr

Benchmarked 2026-07-10
SparkBench · GB10 · single node