CASE STUDY / BLUEPRINT
A complete 75-second technical story—shown beside the exact direction that built it.
Make it technical without becoming a spec sheet. Use evidence. Admit the loss. End on what the efficiency unlocks.
WHY IT’S BUILT THIS WAY
The first draft was not a timeline. It was a point of view: the headline is not that a small model wins every test—it is that six billion active parameters can compete at all.
That decision controls the whole piece. We open on the impossible claim, inspect the architecture, convert the mechanism into a visual metaphor, and then stack enough proof to make the claim credible.
The loss is intentionally kept in. MotionAgent is directing persuasion, not manufacturing hype. That honest beat earns the right to end on the real implication: where capable models can run.
The Blueprint pack supplies the type system, grid, annotation language, chart grammar, palette, and pacing. The agent still has to decide what each scene is for.
SCENE BY SCENE
Eleven instructions. One continuous argument. Every preview below is the exact segment produced by that decision.
Start with the tension, not the setup. The parameter count and the Opus comparison arrive in the same frame.
DIRECTOR INSTRUCTIONOpen with BlueprintTitle. Set the kicker to “Parameter efficiency,” make “Six billion active parameters” the dominant type, and land the Opus comparison as the lede.
Strip away everything except the surprising number and the job it supposedly does.
DIRECTOR INSTRUCTIONUse BlueprintSection with 6B as the hero object. Pair it with “Beats Claude Opus / on agentic work” and keep the benchmark qualifier visible.
Move from claim to source material. The visual proof should feel inspected, annotated, and technical.
DIRECTOR INSTRUCTIONStage the real architecture image inside BlueprintPlate. Annotate QSA, GDN layers, and MoE on every block; keep the source URL in the plate caption.
The explanation needs one image the viewer can remember: hundreds of experts, only a few active.
DIRECTOR INSTRUCTIONBuild a 512-cell BlueprintDotMatrix. Highlight ten routed experts, then label the shared expert separately so the efficiency mechanism reads instantly.
A comparison chart turns the architecture detail into a competitive advantage.
DIRECTOR INSTRUCTIONChart active parameters per token: Flash-Next at 6B, DeepSeek V4 at 13B, and Qwen 3.7 Plus at 17B. Use the hot accent only on Flash-Next.
Three comparisons create a pattern; one comparison could be dismissed as a cherry-picked win.
DIRECTOR INSTRUCTIONRun three compact BlueprintBarChart beats: SWE-Pro 62.5 vs 53.4, JobBench 55.7 vs 36.6, and IFBench 81.3 vs 62.5.
Honest contrast earns trust and makes the efficiency argument stronger, not weaker.
DIRECTOR INSTRUCTIONUse BlueprintCompare for Humanity’s Last Exam: 35.9 for Flash and 40.0 for Opus. Add “That is fair” and frame the small gap as the insight.
The story pivots from leaderboard obsession to the cost of competing at all.
DIRECTOR INSTRUCTIONReturn to BlueprintSection. Lead with a single arrow and the line “Not that it wins / what it costs.” Let the pause do the work.
Context length and model family turn the abstract efficiency claim into something builders can evaluate.
DIRECTOR INSTRUCTIONCreate a BlueprintKpiGrid for 262K native context, extensible to 1M, and the qwen4_exp architecture tag.
One final source plate closes the evidence loop before the conclusion.
DIRECTOR INSTRUCTIONPut the evaluation screenshot in BlueprintPlate. Pull out GPQA 91.7 and SWE-Bench Pro 62.5, then state “Small enough to run anywhere.”
End on what changes for the viewer, not another specification.
DIRECTOR INSTRUCTIONClose with BlueprintQuote: “Six billion active parameters changes where these models can run.” Finish with the “Running it local?” response cue.
THE DIFFERENCE IS DIRECTION