Local Model Benchmarks

Local test machine: Apple M5 Pro · 64 GB unified memoryRun window 29 September — 18 October 202622 recorded runs

29 September 2026 · Agentic scene · voxel garden

Voxel Pagoda Garden

Session completed Local run

Completed the voxel garden challenge and produced a single-file HTML artifact.

Voxel Pagoda Garden with the pagoda, cherry trees and orbit controls visible
Voxel Pagoda Garden · gameplay capture

What was checked

  • JavaScript syntax check passed.
  • Headless stub runtime completed five animation frames.
  • The archived HTML is 23045 bytes, matching the artifact size in the source record.

Configuration

The model, serving engine and agent used for this execution.

Recorded configuration and setup details
FieldValueDetail
Model · primaryQwen 3.8 Flash Next · primaryEXL3 2.625 bpw
RuntimeSushi 1.0.5inference engine and exact build
HarnessPi coding agentthe agent or client that drove the run
HardwareApple M5 Pro · 64 GB unified memoryhardware recorded for this execution
WorkloadVoxel Pagoda Garden · revision 1Single-file browser voxel scene generated through a Pi coding-agent session. Separate from the 10-task suite.
Context window131,072 tokensmaximum context configured for the serving runtime
Repetitions1executions recorded for this workload; one means an indicative result, not an estimate

Inference measurements

Engine-side rates for this run. Each figure states its own statistic; figures from different runs are not interchangeable.

Inference measurements: recorded value and measurement definition
MetricValueDefinition
Cache hit ratio90.23 %cached input tokens / presented input tokens · n=1
Decode · output-weighted44.63 tok/soutput-token weighted mean · n=13
Prefill · median334.2 tok/smedian across requests · n=13
MTP acceptance72.17 %mean across requests · n=13

Agent session counters

Agent-wide usage totals. These include reasoning, tool calls and waiting, and are not model decode speed.

Agent session counters: recorded value and measurement definition
MetricValueDefinition
Session elapsed9.5 minsession start to final delivery · n=1
Model requests13 requeststotal · n=1
Tool calls12 callstotal · n=1
Context compactions0 compactionstotal · n=1

Context and token counters

Token totals as reported by the harness or engine. Cached and cumulative figures are not additive with fresh input.

Context and token counters: recorded value and measurement definition
MetricValueDefinition
Final prompt context20,269 tokensfinal request · n=1

Recorded check counters

Counts from the checks recorded with this run.

Recorded check counters: recorded value and measurement definition
MetricValueDefinition
Reported voxels23,896 voxelsartifact report · n=1
Reported InstancedMesh instances71 meshesartifact report · n=1
Reported scene objects77 objectsartifact report · n=1
Request records & provenance

Numeric per-request measurements only. Stop reasons and all prompt or response content are excluded.

Run ID: 2026-09-29-qwen38-flash-next-sushi-pi-pagoda-garden

One execution, with checks transcribed from the retained record. A completed session is not a quality score. Measurement method · All runs