Skip to content

Dark Scarlett v0.3 26B A4B

Every measured variant of this model: how much quality each quant keeps, and the VRAM and speed it costs to run.

VRAM footprint vs Local Intelligence Index

LB-2026-07.2 | 25/22.5/22.5/22.5/7.5

This chart appears after the model's first measured run lands. Use the quant ladder below for file size and VRAM requirements.

Variant profiles

Complete rows are ordered by Local Intelligence Index; partial rows show their measured axes but are not ranked. The VRAM/Fits columns (8Kcontext) tell you what your card needs. Ranks are within this family's variants.

Column guide

Variant — quant label plus, where sha-verified, the publisher of the exact weights file benchmarked; rows without a source line predate artifact identity records.

VRAM @8k — model weights, KV cache, and runtime headroom at 8k context.

Fits — the smallest common GPU VRAM tier above that estimate.

Prefill tok/s — prompt-processing speed.

Decode tok/s — generated-token speed after the prompt.

Overall tok/s — completion throughput across the full benchmark.

File size — the benchmarked model artifact on disk.

Runtime — the serving engine and version.

Run — the immutable benchmark receipt. Live rows link to the public submission record until the run is baked into the static site.

Pick the largest quant whose VRAM @8k fits your card.

Swipe horizontally for all variant metrics →

Rank (this family)VariantLocal Intelligence Indexindex-v3.0 | 40/15/15/10/15/5VRAM @8kFitsOverall tok/sFile sizeRuntimeRun
no run yet29.2 GB32 GB26.9 GBbenchmark it
no run yet24.9 GB32 GB22.6 GBbenchmark it
no run yet21.4 GB24 GB19.1 GBbenchmark it
no run yet19.1 GB24 GB16.8 GBbenchmark it

vs base

Fine-tune comparison

Dark Scarlett v0.3 26B A4B

Fine-tune of Gemma 4 26B A4B IT

composite n/acompare to base
base not yet benchmarkedfine-tune not yet benchmarked

Measured axis deltas appear after both rows have board data.