viraldi.
Video details
How it works
Back to discovery
17:29
Technology & Computing / VIDEO

The New Way to Run 125B Models 6× Faster Than llama.cpp (Strata)

Kai· 58.4K subscribers

22d ago · English audio

Qwen 3.8 Flash-Next is a 125B model, and a new local AI engine called Strata runs it at 93 tokens a second on a 12GB RTX 5070. The "6× faster than llama.cpp" claim is closer to 2× like for like, and the real cost of running it on a gaming PC is 64GB of RAM. In this video: - Why a 125B MoE model only reads about 4GB per token (6B active) - How Strata's expert offloading splits every layer between GPU and CPU…

Watch on YouTube
5.0×Above the channel median
100.7KViews
0Avg. views / hour · 1.1h window
20KMedian · 50 videos

Outlier score

100.7K views ÷ a 20K baseline = 5.0×. Compared with 50 earlier videos from the same channel, matching the format recorded at scoring time. Older-upload baseline: reference videos had at least as long to gain views. This comparison is not age-matched.

YouTube snapshot: Sun, 25 Oct 2026 12:33:31 GMT. Counters may have changed.

Read the methodology

Generative AI & Agents

Channel-balanced percentile: 96

Comparison details

39 videos from 29 other channels, with each channel weighted equally. Same format and language, with 10,000–99,999 subscribers. Observed at 549 hours after publication, within ±48 hours.

This describes the collected sample, not all of YouTube. Updated Sun, 25 Oct 2026 13:31:01 GMT.

Niches & tags

Transcripts and detailed analytics are available for recent outlier Shorts up to three minutes. Research tools

Breakouts from other channels