commonllama benchmark · model

Qwen3.5 4B

Every number on this page is an observed measurement from a benchmark run, shown with the run that produced it.

Pre-release: these numbers come from commonllama harness 0.0.0, a build not yet released. They are real measurements, not estimates; released-harness numbers will replace them.

Observed facts

What the harness read from the file.

Model file digestc6f6972469c72b8081f3e5b07c1f24379330fccfc2f7f4fd3b70e330d98465ec
Architecturenot observed
Quantizationnot observed
Max contextnot observed
Model file size2.52 GiB (2,707,514,208 bytes)
Fact provenanceunknown

Context capacity

How much context this model has carried.

Qualified to at least 131,072 tokens of context, an observed lower bound, not a maximum.

Observed on goldspier-rtx5090 vulkan. A frontier is a property of one run on one machine class, never a promise about another.

Saved context

What a saved context costs on disk.

Each curve keeps to one named key: this model, one backend. Unlike measurements are never folded together.

The KV cache is the AI’s memory of the conversation so far. Saving it to disk lets the AI return to the conversation and load that memory back in, rather than read everything from the start again.

Bytes per cached token, complete runs only
BackendContextBytes per tokenSaved bytes
cuda4,096unavailableunavailable
cuda8,192unavailableunavailable
cuda16,384unavailableunavailable
cuda32,768unavailableunavailable
cuda65,536unavailableunavailable
cuda131,072unavailableunavailable
cuda4,096unavailableunavailable
cuda8,192unavailableunavailable
cuda16,384unavailableunavailable
cuda32,768unavailableunavailable
cuda65,536unavailableunavailable
cuda131,072unavailableunavailable
vulkan4,096unavailableunavailable
vulkan8,192unavailableunavailable
vulkan16,384unavailableunavailable
vulkan32,768unavailableunavailable
vulkan65,536unavailableunavailable
vulkan131,072unavailableunavailable
vulkan4,096unavailableunavailable
vulkan8,192unavailableunavailable
vulkan16,384unavailableunavailable
vulkan32,768unavailableunavailable
vulkan65,536unavailableunavailable
vulkan131,072unavailableunavailable

Provenance

Every run behind the numbers.

Published runs
RunProvenanceStatusHardware classRAM bucketBackendStorageFrontier
0032b87abd93migrationcompletegoldspier-rtx5090unavailablevulkanssd131,072+
6cb47bc304bfmigrationcompletegoldspier-rtx5090unavailablecudassd131,072+

A partial run is content, not a failure: its rungs carry typed reasons on the run page. Partial and complete runs are never mixed inside one aggregate on this page.

commonllama benchmark records are dedicated to the public domain under CC0 1.0. Use them for anything.