Resolve + local AI benchmarks

Version-to-version measurements of a fully local DaVinci Resolve Studio + Splash LLM workstation. Every number comes from an independent read-back. See the stack overview if you are new.

Scorecards

Benchmark reruns, all four suitesMCP 5.2.2 · Splash 1.6.1
Agentic 18/18 (sixth clean wave), Splash plateau +39% on INT8 with the peak moving to width 12, agentic suite 12/12, OpenCV bit-identical — with the review-agent corrections folded in.
Splash 1.6.0 reproducibility rerun1.6.0
Same-build rerun establishing the ~2% spread band and reproducing every ladder-shape feature, including the width-6 dip.
MCP 5.2.0 agentic wave21.2.0.6 · MCP 5.2.0
18/18 fifth wave; append_to_timeline not pin-able — real cross-writes caught, reversed and disclosed by the agents themselves.
MCP 5.1.0 agentic wave21.2.0.6 · MCP 5.1.0
First 5.1.0 pass: 18/18, the record_frame trio, and append moving to media_pool.
Splash engine baseline1.6.0
TTFT, the width-ladder plateau at 372–383 tok/s, and the KV-pressure chase for a first suspension that never came.
MCP 4.19.3 re-measure21.2.0.6 · MCP 4.19.3
The replacement pass: two withdrawn rows, a single-tenant probe at 1191 ms, and the render tax measured from inside a two-agent wave.
Resolve MCP benchmark21.2.0.6 · MCP 4.19.0
Render probes, server-side operation medians, and the six-task agentic suite run by sealed fresh agents.
Resolve MCP benchmark21.1.2.10 · MCP 4.18.5
The baseline run: cold/warm renders, agentic n=3, concurrency sweep.

Documentation & data