YouTube Video
Want to make money and save time with AI? Join here: https://www.skool.com/ai-profit-lab-7462/about
Video notes + links to the tools 👉 https://www.skool.com/ai-profit-lab-7462/about
Get a FREE AI Course + Community + 1,000 AI Agents 👉 https://www.skool.com/ai-seo-with-julian-goldie-1553/about
Get a FREE AI SEO Strategy Session: https://go.juliangoldie.com/strategy-session?utm=julian
Tencent shrunk its 770B-parameter HY4 Preview from 1.5TB to 213GB — but the number everyone's leaving out of the thumbnail changes the whole story. I break down exactly how the mixed-precision quantization works, what it actually scores on benchmarks, and the one line in the docs that stops most people from running it. If you care about open models and where local AI is really headed, this is the honest version.
00:00 Intro – 1.5TB to 200GB, and the missing number
00:40 The Real Math – Why HY4 Preview is 1.56TB
01:06 213GB – The compressed build
01:14 Quantization Explained – In plain English
01:39 The Smart Trick – Mixed precision, not brute force
02:00 The Photo Analogy – Why some layers get protected
02:09 Where the Cuts Go – Experts vs down projection
02:27 Benchmark Results – 85% smaller, ~1 point lost
03:38 The Catch – You still need 214GB of VRAM
03:57 The Better Option – Why Q4_K_M is the default
04:16 Storage vs VRAM – The mistake everyone makes
04:24 Real Speed – 20 tokens/sec on 8 data center GPUs
04:39 Who It's Actually For – Not your desktop
05:19 Inside the Model – MoE, 49B active per token
05:49 Sparse Attention & 1M Context – The efficiency stack
06:06 It Won't Just Run – Stock llama.cpp fails
06:41 The Bigger Pattern – 4 techniques, one direction
07:20 The Real Takeaway – Spend precision where it matters
Video notes + links to the tools 👉 https://www.skool.com/ai-profit-lab-7462/about
Get a FREE AI Course + Community + 1,000 AI Agents 👉 https://www.skool.com/ai-seo-with-julian-goldie-1553/about
Get a FREE AI SEO Strategy Session: https://go.juliangoldie.com/strategy-session?utm=julian
Tencent shrunk its 770B-parameter HY4 Preview from 1.5TB to 213GB — but the number everyone's leaving out of the thumbnail changes the whole story. I break down exactly how the mixed-precision quantization works, what it actually scores on benchmarks, and the one line in the docs that stops most people from running it. If you care about open models and where local AI is really headed, this is the honest version.
00:00 Intro – 1.5TB to 200GB, and the missing number
00:40 The Real Math – Why HY4 Preview is 1.56TB
01:06 213GB – The compressed build
01:14 Quantization Explained – In plain English
01:39 The Smart Trick – Mixed precision, not brute force
02:00 The Photo Analogy – Why some layers get protected
02:09 Where the Cuts Go – Experts vs down projection
02:27 Benchmark Results – 85% smaller, ~1 point lost
03:38 The Catch – You still need 214GB of VRAM
03:57 The Better Option – Why Q4_K_M is the default
04:16 Storage vs VRAM – The mistake everyone makes
04:24 Real Speed – 20 tokens/sec on 8 data center GPUs
04:39 Who It's Actually For – Not your desktop
05:19 Inside the Model – MoE, 49B active per token
05:49 Sparse Attention & 1M Context – The efficiency stack
06:06 It Won't Just Run – Stock llama.cpp fails
06:41 The Bigger Pattern – 4 techniques, one direction
07:20 The Real Takeaway – Spend precision where it matters
SOCIAL SHARE CARD GENERATOR