← deepseek-v4-preview   ·   Motion capture page (noindex)

Motion Capture

DeepSeek V4 Preview Matches Frontier Code at One-Sixth the Price

Recording Mode Off — plain dark background
Lower Third Name Tag

Guest or host ID. Slides up, holds, exits at 3s.

Jane Sterling
AI Research Lead
Sterling Intelligence

Outro CTA End Screen

7-second closer. YouTube end screen cards go over the two bottom corners — add them in YouTube Studio after upload.

If the receipts hit —

Subscribe.

New episodes every week. No filler.

YouTube@SterlingIntelligence
X / Twitter@SterlingIntel
WebSterlingintel.ca

Stat Callout Card Data Overlay

Slides in over footage to highlight a key number. Counts up automatically. Auto-dismisses at 3s.

DeepSeek V3 Debut
0mo
Triggered sharp selloff in US AI & semiconductor stocks
April 2025 — forced reassessment of compute assumptions globally
V4 repeats the pattern

Multi-Metric Strip Dashboard

Three stats in a bottom bar that slides up over footage.

DeepSeek V4-Pro Specs
1.6T
Total Parameters
49B
Active Per Inference
1M
Token Context Window

Model Comparison Card Head-to-Head

Side-by-side benchmark scores with animated count-up.

SWE-bench Verified
DeepSeek V4-Pro
0%
VS
Claude Opus 4.6
0%

Model Comparison Card Head-to-Head

Side-by-side benchmark scores with animated count-up.

Codeforces Rating
DeepSeek V4-Pro
0pts
VS
GPT-5.4
0pts

Multi-Metric Strip Dashboard

Three stats in a bottom bar that slides up over footage.

V4-Pro Benchmark Sweep
90.1%
GPQA Diamond
87.5%
MMLU-Pro
95.2%
HMMT 2026 Math

Stat Callout Card Data Overlay

Slides in over footage to highlight a key number. Counts up automatically. Auto-dismisses at 3s.

MRCR 1M Retrieval
0%
Long-context needle-in-haystack test at full 1M token depth
DeepSeek V4-Pro · April 2026
Context window is real

Model Comparison Card Head-to-Head

Side-by-side benchmark scores with animated count-up.

Terminal-Bench 2.0
DeepSeek V4-Pro
0%
VS
GPT-5.5
0%

Stat Callout Card Data Overlay

Slides in over footage to highlight a key number. Counts up automatically. Auto-dismisses at 3s.

SWE-bench Pro
0%
Harder real-world coding agent benchmark — complex multi-step tasks
DeepSeek V4-Pro · April 2026
Gap widens on harder eval

Multi-Metric Strip Dashboard

Three stats in a bottom bar that slides up over footage.

Hybrid Attention Efficiency
27%
Inference FLOPs vs V3.2
10%
KV Cache vs V3.2
32T+
Training Tokens

Model Comparison Card Head-to-Head

Side-by-side benchmark scores with animated count-up.

API Input Price (per M tokens)
DeepSeek V4-Pro
0$
VS
GPT-5.5
0$

Stat Callout Card Data Overlay

Slides in over footage to highlight a key number. Counts up automatically. Auto-dismisses at 3s.

V4-Flash Input Price
0
$0.14 per million input tokens — among the cheapest capable AI APIs
DeepSeek V4-Flash · April 2026
36× cheaper than GPT-5.5 input

Finance Chart Card Comparative Performance

Two-ticker line chart fetched live from Yahoo Finance. Normalized to 100 at period start so both lines start from the same baseline. Animated draw-in over 3s.

Nvidia vs SMIC — Last 3 Months
Since DeepSeek V4's Huawei chip optimization signal
Loading data…
Nvidia SMIC Source: Yahoo Finance · April 2026

Growth Rate Card YoY / MoM

Percentage growth with animated count-up. Slides in from the right, exits at 3s.

SMIC Stock Surge
+ 0%
SMIC Hong Kong shares
V4 Launch Day · April 24, 2026

Stat Callout Card Data Overlay

Slides in over footage to highlight a key number. Counts up automatically. Auto-dismisses at 3s.

Fake Accounts (Anthropic Probe)
0
Accounts traced to Chinese AI firms accessing Claude API without authorization
Anthropic Security Investigation · 2026
DeepSeek among firms named

Stat Callout Card Data Overlay

Slides in over footage to highlight a key number. Counts up automatically. Auto-dismisses at 3s.

Stolen API Queries
0M
Unauthorized Claude API queries used in apparent model distillation operations
Anthropic Security Investigation · 2026
FMF coalition formed in response
Capture order: top to bottom. Click any preview to enter fullscreen with a 3-second countdown — perfect for OBS.
Recording: use OBS Window Capture cropped to the preview frame, or fullscreen + display capture. Hit Replay, then record. Output drops as an overlay track in your editor.
Palette: the palette saved in motion_kit auto-applies here via localStorage.
Editor workflow — what to do after capturing here:
  1. Generate the avatar look in HeyGen using this episode's HEYGEN_LOOK prompt + the SCRIPT field as the read. Kick the render and grab a coffee.
  2. While HeyGen renders: capture the motion graphics above (top-to-bottom = capture order), grab b-roll, gather logos / photos as transparent PNGs, drop everything into your working folder.
  3. When HeyGen finishes: download both the rendered video and the timecoded transcript (.srt or .vtt). Drop the transcript at:
    /home/jaysoncraig/public_html/sandbox/data/faces/.build/deepseek-v4-preview/transcript.srt
  4. Generate the Resolve marker CSV + printable shot-list:
    python3 /home/jaysoncraig/.claude/scripts/generate-resolve-markers.py --slug deepseek-v4-preview
    Outputs: /home/jaysoncraig/public_html/sandbox/data/faces/.build/deepseek-v4-preview/timeline-markers.csv and /home/jaysoncraig/public_html/sandbox/data/faces/.build/deepseek-v4-preview/shot-list.html
  5. Open DaVinci Resolve, set the project framerate (the marker tool emits 30fps by default; pass --framerate 60 if your project is 60), drop the HeyGen video on V1.
  6. Timeline header → Import → Marker List from CSV → select timeline-markers.csv. Every cue lands as a colored marker on the correct track: Yellow=TalkingHead, Cyan=VoiceoverScene, Red=Cut, Green=B-roll, Pink=StatCard.
  7. Drop motion-graphic captures, b-roll clips, and stat-card overlays onto the marked positions. The shot-list HTML is your printable fallback if any markers look off.
  8. Export → render → upload. After publish, drop the YouTube URL into the episode's .repurpose.txt placeholders for the cross-channel pack.