<?xml version="1.0" encoding="UTF-8"?>
<feed xmlns:yt="http://www.youtube.com/xml/schemas/2015" xmlns:media="http://search.yahoo.com/mrss/" xmlns="http://www.w3.org/2005/Atom">
 <link rel="self" href="http://www.youtube.com/feeds/videos.xml?channel_id=UCvN3QmpWy_eFG8gIU_Ij-RQ"/>
 <id>yt:channel:vN3QmpWy_eFG8gIU_Ij-RQ</id>
 <yt:channelId>vN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
 <title>Zero to MVP</title>
 <link rel="alternate" href="https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ"/>
 <author>
  <name>Zero to MVP</name>
  <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
 </author>
 <published>2025-06-30T15:10:41+00:00</published>
 <entry>
  <id>yt:video:Jw99x0S7s_4</id>
  <yt:videoId>Jw99x0S7s_4</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>Why I Run 4B and 125B AI Models 24/7</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=Jw99x0S7s_4"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-11-16T16:51:20+00:00</published>
  <updated>2026-11-18T17:36:36+00:00</updated>
  <media:group>
   <media:title>Why I Run 4B and 125B AI Models 24/7</media:title>
   <media:content url="https://www.youtube.com/v/Jw99x0S7s_4?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i3.ytimg.com/vi/Jw99x0S7s_4/hqdefault.jpg" width="480" height="360"/>
   <media:description>Three months ago I showed the tiny 2–4 GB local AI models I keep running all the time. Now bigger ones run right next to them, including a 125-billion-parameter model for code, all on one always-on home server with a 16 GB graphics card.

In this video I go through my current local AI setup: which model I use for each task, how fast it runs, and why I still use small specialized models alongside the big one.

Models in this video:
• LightOnOCR 3 (4B): documents, screenshots, tables, formulas, handwriting
• Gemma 4 (26B): article summaries, about 70 tokens per second
• TranslateGemma (12B): translation across 55 languages
• MedGemma (27B): medical information for me and my family
• Qwen 3.8 Flash-Next (125B): code, about 50 tokens per second

Hardware: XAIStack, a modular mini PC with a NAS block and a GPU block, Intel Core Ultra 5,  RTX 4060 Ti 16 GB. About 40 W at idle with the graphics card installed.

🔗 Links:
XAIStack on Kickstarter: https://www.kickstarter.com/projects/xaistack/xaistack?ref=8pa1e3
Previous video, tiny models I run all the time: https://youtu.be/9LkxI2H3gI0
How to run a big model on regular graphics cards: https://youtu.be/B77LasklmZU
My Newsletter: https://weekly.blokhin.us
GridCore: https://github.com/w512/GridCore

⏱️ Timestamps
00:00 From tiny models to 125B
00:31 Why I use local models
01:21 The device: XAIStack
02:53 What &quot;running all the time&quot; means
03:18 OCR
03:56 Summaries
04:29 Translation
05:00 Medical documents
05:24 Code
06:23 The right model for each task
07:16 Who XAIStack is for


#LocalAI #LocalLLM #Homelab</media:description>
   <media:community>
    <media:starRating count="560" average="5.00" min="1" max="5"/>
    <media:statistics views="52082"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:xHesuFZw_Gc</id>
  <yt:videoId>xHesuFZw_Gc</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>How to Make AI Videos From Scratch: 3 Prompts, 10 Minutes</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=xHesuFZw_Gc"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-11-12T17:34:14+00:00</published>
  <updated>2026-11-19T06:25:53+00:00</updated>
  <media:group>
   <media:title>How to Make AI Videos From Scratch: 3 Prompts, 10 Minutes</media:title>
   <media:content url="https://www.youtube.com/v/xHesuFZw_Gc?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i1.ytimg.com/vi/xHesuFZw_Gc/hqdefault.jpg" width="480" height="360"/>
   <media:description>New users get up to 50% off their first purchase on Higgsfield 👉  https://higgsfield.ai/s/make-your-first-ai-video-yt-zerotomvp-ezpjDv

In this Higgsfield tutorial I make an AI video from scratch on Higgsfield AI: no camera, no actor, no experience needed, about 10 minutes and 215 credits. This is how to make AI videos as a complete beginner. I create a character with GPT Image 2.5, build a character sheet so he stays the same in every shot, and turn it into a realistic AI video with sound in Seedance 2.5. Then I change one line of the prompt to get a second version, and as a bonus put the same character into a trending video with Genjutsu. Every model lives in one AI video generator, on one plan, so AI video for beginners really is three prompts. Try it yourself, the prompts are below. This video is sponsored by Higgsfield.

-------------------------------------------------------

📝 My prompts

Character (GPT Image 2.5 Flare, 3:4):
A realistic full-body photo of a 65-year-old software engineer with a short grey beard, messy grey hair and thin round wire-rimmed glasses, wearing a dark-green knitted cardigan over a blue checked flannel shirt, brown corduroy trousers and worn leather slippers. He holds a chunky old grey laptop with no logos under one arm. He stands naturally against a plain light-grey background, facing the camera, face clearly visible. Soft, even studio lighting, natural skin texture, realistic proportions, sharp focus.

Character sheet (GPT Image 2.5 Sunburst, 16:9):
Use the attached image as the only reference. Create a character sheet of this exact man in three panels side by side: a full-body front view, a full-body side view facing right, and a close-up of his face. Keep the same face, beard, glasses, hair, cardigan, shirt, trousers, slippers and laptop in every panel, with identical colours and patterns. Plain light-grey background, even studio lighting, realistic photo. No text, labels or borders.

Video (Seedance 2.5, 15 s, 16:9, 720p):
@Image1 is the main character. Late at night he sits at a cluttered wooden desk in a small dark room, typing fast on the old grey laptop. The screen lights his face with a cold glow, and rain runs down the window behind him. He stops, frowns at the screen, types one more line, then leans back and smiles as if the code finally works. Keep his face, beard, glasses and cardigan exactly the same throughout. One continuous shot, the camera slowly pushing in from the side. Sound: rain on the window, loud mechanical keyboard clicks, a quiet satisfied laugh at the end. No music, no dialogue, no text on screen.

-------------------------------------------------------

⏱ Chapters
00:00 What can a beginner make with AI video today?
00:51 Which AI model should a beginner use?
02:55 How do you keep the same character with GPT Image 2.5?
03:31 How do you make a video with Seedance 2.5?
05:30 How do you put your character into a trend with Genjutsu?
06:08 Conclusions


#AIvideo #HiggsfieldAI #Seedance25 #AIvideogenerator #AIforbeginners</media:description>
   <media:community>
    <media:starRating count="17" average="5.00" min="1" max="5"/>
    <media:statistics views="14758"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:wLN8QeUNFhc</id>
  <yt:videoId>wLN8QeUNFhc</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>Fable 5.2 — a routine update or a real step forward?</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=wLN8QeUNFhc"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-11-10T17:34:17+00:00</published>
  <updated>2026-11-18T09:29:34+00:00</updated>
  <media:group>
   <media:title>Fable 5.2 — a routine update or a real step forward?</media:title>
   <media:content url="https://www.youtube.com/v/wLN8QeUNFhc?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i4.ytimg.com/vi/wLN8QeUNFhc/hqdefault.jpg" width="480" height="360"/>
   <media:description>Today's partner:
Junie Lite – free AI layer, no trial, no card: https://jb.gg/ZeroToMVP


Fable 5.2 is out, and Anthropic says it beats both Fable 5.1 and Opus 5.5 on almost every benchmark it lists. Instead of running synthetic tests, I gave Fable 5.2 and Opus 5.5 the same code review prompt on a real project: my scheduler for local AI models, about 20,000 lines of Go.

---------------------------------

My Weekly Newsletter: https://weekly.blokhin.us
Patreon: https://patreon.com/ZeroToMVP


⏱️ Timestamps
00:00 Fable 5.2 is out
01:00 Junie Lite
01:35 What's new in Fable 5.2
02:41 My thoughts
03:11 Quick test: code review on 20K lines of Go
04:00 Results: Fable 5.2 vs Opus 5.5
04:35 Is it worth paying double?


#Fable52 #ClaudeOpus #AICoding</media:description>
   <media:community>
    <media:starRating count="49" average="5.00" min="1" max="5"/>
    <media:statistics views="20907"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:Abgo6kY-4wY</id>
  <yt:videoId>Abgo6kY-4wY</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>I Tried to Break DeepSeek V4.1 Flash on 6 Backend Tasks</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=Abgo6kY-4wY"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-11-09T15:33:00+00:00</published>
  <updated>2026-11-16T01:22:58+00:00</updated>
  <media:group>
   <media:title>I Tried to Break DeepSeek V4.1 Flash on 6 Backend Tasks</media:title>
   <media:content url="https://www.youtube.com/v/Abgo6kY-4wY?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/Abgo6kY-4wY/hqdefault.jpg" width="480" height="360"/>
   <media:description>I gave DeepSeek V4.1 Flash six real backend tasks, three runs each: permissions, a data migration, a slow page, double votes and file storage. Every counted run passed all hidden tests, and all 18 together cost about 66 cents. In one extra run, the model found my reference solutions on disk and copied them.

This video is sponsored by Neon. Try Neon for free: get.neon.com/VQPkVEm

---------------------------------------

My Weekly newsletter:  https://weekly.blokhin.us
Previous video, DeepSeek V4.1 Flash vs Claude Opus 5.5: https://youtu.be/NddJHzLfSf4

Which model should go through this ladder next? Let me know in the comments.

---------------------------------------

⏱️ Timestamps
00:00 Intro
01:03 Test setup
01:55 How the test works
03:47 Task 1: Editing posts
04:05 Task 2: Private boards
04:37 Task 3: Migration without data loss
05:23 Task 4: The slow page
05:55 Task 5: Double votes
06:25 Task 6: Screenshots, and the run that found my answers
07:55 What it cost
08:27 Takeaways



#DeepSeek #AICoding #AgenticCoding</media:description>
   <media:community>
    <media:starRating count="88" average="5.00" min="1" max="5"/>
    <media:statistics views="22032"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:S44fqK0FI84</id>
  <yt:videoId>S44fqK0FI84</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>Ripple 65B: A Chip Lab Beat Every US Open Model. I Tested It</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=S44fqK0FI84"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-11-06T15:41:21+00:00</published>
  <updated>2026-11-09T23:41:35+00:00</updated>
  <media:group>
   <media:title>Ripple 65B: A Chip Lab Beat Every US Open Model. I Tested It</media:title>
   <media:content url="https://www.youtube.com/v/S44fqK0FI84?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i4.ytimg.com/vi/S44fqK0FI84/hqdefault.jpg" width="480" height="360"/>
   <media:description>Ripple-65B Preview is a new open coding model from Ricursive Intelligence, a chip design company founded by the authors of Google's AlphaChip. It scores 80.5 on Terminal-Bench 2.1 and is positioned as the best US open model for agentic coding. I ran it locally on a single 32 GB GPU with llama.cpp to see whether it can handle real tasks.

🖥️  Setup:
• Model: Ripple-65B Preview, 26 GB quantization
• Runtime: llama.cpp built from an open pull request, 128K context
• Agent: Pi, sandbox without internet, hidden checks
• GPU: AMD Radeon Pro W7800, 32 GB VRAM – https://link.amazon/B00A7pMwl
• CPU: Ryzen 7 – https://link.amazon/B0gBJk64B
• RAM:128 GB – https://link.amazon/B0079ZOHt


📬 My Weekly Newsletter — https://weekly.blokhin.us


⏱️ Timestamps
00:00 A coding model from a chip design lab
00:54 The benchmark table and the MoE catch
02:30 Running it on a 32 GB GPU
03:22 A first small task
03:45 Terminal tasks: C build, git, Python service
04:48 A big project: tower defense
05:21 Conclusions



#LocalLLM #AICoding #OpenSourceAI 

P.S. As an Amazon Associate I earn from qualifying purchases.</media:description>
   <media:community>
    <media:starRating count="101" average="5.00" min="1" max="5"/>
    <media:statistics views="9115"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:KrWBNIaQQD8</id>
  <yt:videoId>KrWBNIaQQD8</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>15 Cents a Run: Are Chinese Flash Models Good Enough to Code?</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=KrWBNIaQQD8"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-11-03T16:36:14+00:00</published>
  <updated>2026-11-13T09:05:11+00:00</updated>
  <media:group>
   <media:title>15 Cents a Run: Are Chinese Flash Models Good Enough to Code?</media:title>
   <media:content url="https://www.youtube.com/v/KrWBNIaQQD8?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i4.ytimg.com/vi/KrWBNIaQQD8/hqdefault.jpg" width="480" height="360"/>
   <media:description>Four Chinese &quot;flash&quot; models (DeepSeek, MiniMax, GLM and MiMo) cost tens of times less than the top AI models. Are they good enough for real coding work? I gave each one two very different jobs, ran each job twice per model, and graded every run automatically.

---------------------------------

Ad: This video is sponsored by TripleTen, a professional education platform designed to help tech professionals and career changers move into higher-paying tech and AI roles.

Explore two TripleTen programs for experienced engineers:

AI Systems Engineering (22 weeks)
https://tripleten.com/systems-engineer/?utm_source=youtube&amp;utm_medium=video&amp;utm_campaign=zerotomvp

AI and Machine Learning (20 weeks)
https://tripleten.com/ai-machine-learning/?utm_source=youtube&amp;utm_medium=video&amp;utm_campaign=zerotomvp

Both are for engineers with 2+ years of recent, hands-on production experience and require a short skills audit before you can enroll. Both also include career coaching and mentorship to help you stand out in the job market.

---------------------------------

⏱️ Timestamps
00:00 Four cheap models, two real jobs
01:30 Task 1: a tower defense game from a spec
02:24 Game results
03:37 Sponsor: TripleTen
05:18 Task 2: a security audit with 13 vulnerabilities
06:02 Security results
07:25 Which model for which job




#DeepSeek #AICoding #LLM</media:description>
   <media:community>
    <media:starRating count="182" average="5.00" min="1" max="5"/>
    <media:statistics views="20064"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:WbvIlrqOQ-o</id>
  <yt:videoId>WbvIlrqOQ-o</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>I Treated AI Video Like Code. Here’s What Broke</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=WbvIlrqOQ-o"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-10-30T19:52:38+00:00</published>
  <updated>2026-10-31T08:16:27+00:00</updated>
  <media:group>
   <media:title>I Treated AI Video Like Code. Here’s What Broke</media:title>
   <media:content url="https://www.youtube.com/v/WbvIlrqOQ-o?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i4.ytimg.com/vi/WbvIlrqOQ-o/hqdefault.jpg" width="480" height="360"/>
   <media:description>Try Kling 4.0 on Higgsfield: https://higgsfield.ai/s/higgsfield-kling-4-0-yt-zerotomvp-AEmiVf

Kling 4.0 on Higgsfield is an AI video generator that makes a 30 second AI video in one clip. You get three modes in one place: References for multi-reference video, Frames for keyframes AI video and Edit video for AI video editing. I wrote the tests before generating: eight tests, six passed. AI filmmaking you can actually check. 


-------------------------------------------------------

🎬 Kling 4.0 on Higgsfield — https://higgsfield.ai/blog/kling-4-0-on-higgsfield
📬 Newsletter — https://weekly.blokhin.us

-------------------------------------------------------

Chapters:
00:00 Can an AI video pass tests I wrote before generating it?
00:31 What's new in Kling 4.0 on Higgsfield?
00:57 How do you write unit tests for an AI video?
01:41 How do you set up keyframes in Higgsfield Frames mode?
02:41 Did Kling 4.0 hit every keyframe in one 30-second take?
05:02 Can AI cut an eight-frame montage on the marks?
05:05 What happens when the toys come alive?
05:55 Does Kling 4.0 Edit video change only what you ask?
07:15 So is it worth it for planned shots?


#Kling4 #Higgsfield #AIVideoGenerator #AIFilmmaking #AIVideoEditing #AIVideo</media:description>
   <media:community>
    <media:starRating count="23" average="5.00" min="1" max="5"/>
    <media:statistics views="35613"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:w6ZZtYtqW6Q</id>
  <yt:videoId>w6ZZtYtqW6Q</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>How My MacBook Turns Into an AI Workstation</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=w6ZZtYtqW6Q"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-10-28T12:27:29+00:00</published>
  <updated>2026-11-02T00:41:30+00:00</updated>
  <media:group>
   <media:title>How My MacBook Turns Into an AI Workstation</media:title>
   <media:content url="https://www.youtube.com/v/w6ZZtYtqW6Q?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i4.ytimg.com/vi/w6ZZtYtqW6Q/hqdefault.jpg" width="480" height="360"/>
   <media:description>Away from my desk, it's just a MacBook. Plug in one dock, and it's a local AI workstation: two monitors, a NAS, and a Linux desktop with 32 GB of VRAM. Here's how it all fits together.

🖥️  GEAR IN THIS VIDEO
Dock iVANKY FusionDock Ultra — https://bit.ly/4zJSnSx
MacBook M4 Pro — https://link.amazon/B06m8gUo5
Linux desktop: 
    — AMD Ryzen 7: https://link.amazon/B0gBJk64B, 
    — 128 GB RAM: https://link.amazon/B0079ZOHt, 
    — AMD Radeon Pro W7800: https://link.amazon/B00A7pMwl
ZimaCube 2 — https://link.amazon/B0g7VJcXC
Keyboard —  https://link.amazon/B02cu02Uj
Headphones — https://link.amazon/B0aXeBpjr
Microphone — https://link.amazon/B0aex0PJK

------------------------------------------------------

📬 My Weekly Newsletter — https://weekly.blokhin.us


⏱️ Timestamps
00:00 A MacBook that becomes an AI workstation
00:47 One dock for everything
02:03 Two monitors, two browsers
02:37 Linux desktop for local LLMs
03:01 Always-on models on the ZimaCube

#LocalAI #DeskSetup #MacBook 

P.S. As an Amazon Associate I earn from qualifying purchases.</media:description>
   <media:community>
    <media:starRating count="120" average="5.00" min="1" max="5"/>
    <media:statistics views="23146"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:I611wsBQaKQ</id>
  <yt:videoId>I611wsBQaKQ</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>$7 vs $0.65: Opus 5.5 and MiniMax M3.1 Flash on My Go Project</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=I611wsBQaKQ"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-10-27T12:30:35+00:00</published>
  <updated>2026-11-13T09:19:16+00:00</updated>
  <media:group>
   <media:title>$7 vs $0.65: Opus 5.5 and MiniMax M3.1 Flash on My Go Project</media:title>
   <media:content url="https://www.youtube.com/v/I611wsBQaKQ?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/I611wsBQaKQ/hqdefault.jpg" width="480" height="360"/>
   <media:description>Try MiniMax Code Now: https://agent.minimax.io
Download MiniMax Code Desktop: https://agent.minimax.io/download
Hugging Face: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash

I gave MiniMax M3.1 Flash and Claude Opus 5.5 the same three tasks from my real 20,000-line Go project. Both passed every hidden check. One cost about $7, the other $0.65.

------------------------------------------------------

📬 My Weekly Newsletter — https://weekly.blokhin.us
⚙️ GridCore — https://github.com/w512/gridcore


Chapters:
00:00 Intro
01:22 Space Bunny is MiniMax M3.1 Flash
02:04 How I set it up
03:11 Task 1: AMD GPU support
04:45 Task 2: Who's on the GPU
05:54 Task 3: Finding bugs with tests
06:44 Results: time, cost, bugs found
07:33 Where to try MiniMax M3.1 Flash


#MiniMax #ClaudeOpus #AICoding</media:description>
   <media:community>
    <media:starRating count="102" average="5.00" min="1" max="5"/>
    <media:statistics views="21493"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:a1GDbkJwVB8</id>
  <yt:videoId>a1GDbkJwVB8</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>Claude vs DeepSeek vs Local Qwen 3.8: Who Directs a Better AI Video Edit?</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=a1GDbkJwVB8"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-10-21T13:02:06+00:00</published>
  <updated>2026-10-25T20:56:15+00:00</updated>
  <media:group>
   <media:title>Claude vs DeepSeek vs Local Qwen 3.8: Who Directs a Better AI Video Edit?</media:title>
   <media:content url="https://www.youtube.com/v/a1GDbkJwVB8?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/a1GDbkJwVB8/hqdefault.jpg" width="480" height="360"/>
   <media:description>Try Higgsfield Katana: https://higgsfield.ai/s/katana-inside-claude-yt-zerotomvp-ahYkxS

Higgsfield Katana remakes an edit that already works shot for shot, with you in it. It runs as Katana in Claude through the Higgsfield MCP: Claude directs, and Higgsfield generates the shots and does the AI video editing. I gave it one selfie and let Claude Opus, DeepSeek and my local Qwen each direct the same viral edit remake, and not all of them played by the rules. It works for any aura farming edit or AI anime edit, no Claude Motion needed. This video is sponsored by Higgsfield.

-------------------------------------------------------

🗡️ Katana page — https://higgsfield.ai/s/katana-inside-claude-yt-zerotomvp-ahYkxS
📬 Newsletter — https://weekly.blokhin.us

-------------------------------------------------------

⏱️ Timestamps
00:00 Can AI remake a viral edit from one photo?
01:06 What is Higgsfield Katana?
01:46 How do you install Higgsfield in Claude?
02:01 How does Katana in Claude remake a preset?
03:24 Can another model direct Katana in Claude Code?
04:32 Did DeepSeek play by Katana's rules?
05:42 Can a local LLM on a home GPU finish a Katana edit?
06:43 Which AI director is worth using with Katana?


#HiggsfieldKatana #KatanaInClaude #HiggsfieldMCP #AIVideoEditing #Claude #LocalLLM</media:description>
   <media:community>
    <media:starRating count="43" average="5.00" min="1" max="5"/>
    <media:statistics views="19104"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:JFUQAdk7_WY</id>
  <yt:videoId>JFUQAdk7_WY</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>Can an Agent Make a Small Local Model Reliable?</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=JFUQAdk7_WY"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-10-19T16:28:57+00:00</published>
  <updated>2026-10-23T02:55:19+00:00</updated>
  <media:group>
   <media:title>Can an Agent Make a Small Local Model Reliable?</media:title>
   <media:content url="https://www.youtube.com/v/JFUQAdk7_WY?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i3.ytimg.com/vi/JFUQAdk7_WY/hqdefault.jpg" width="480" height="360"/>
   <media:description>Same local model, same GPU, same tasks, two different agents. In one, Qwen 3.8 27B made zero mistakes across nine runs. In the other, it got the numbers wrong twice and confidently reported that everything was correct.

🔗 Links
PilotDeck on GitHub: https://github.com/OpenBMB/PilotDeck
Weekly newsletter: https://weekly.blokhin.us

⏱️ Timestamps
00:00 Same model, two agents
00:45 PilotDeck setup with a local model
02:13 The task: Stripe exports to an investor report
03:10 Level 1: revenue table
03:46 Level 2: MRR and churn
04:54 Level 3: the investor report
06:11 The twist: it wasn't the skills
07:33 The price: time and energy
08:07 Takeaway


#LocalAI #LocalLLM #AIAgents</media:description>
   <media:community>
    <media:starRating count="224" average="5.00" min="1" max="5"/>
    <media:statistics views="21881"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:NddJHzLfSf4</id>
  <yt:videoId>NddJHzLfSf4</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>New DeepSeek V4.1 Flash vs Claude Opus 5.5 on a 20,000-Line Project</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=NddJHzLfSf4"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-10-18T20:48:03+00:00</published>
  <updated>2026-10-22T03:08:31+00:00</updated>
  <media:group>
   <media:title>New DeepSeek V4.1 Flash vs Claude Opus 5.5 on a 20,000-Line Project</media:title>
   <media:content url="https://www.youtube.com/v/NddJHzLfSf4?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i3.ytimg.com/vi/NddJHzLfSf4/hqdefault.jpg" width="480" height="360"/>
   <media:description>DeepSeek V4.1 Flash solved my hardest real-project coding task for 6 cents. Claude Opus 5.5 got the same result for $5.45. But what DeepSeek did while it worked surprised me more than the price.

------------------------------

📬 Weekly newsletter: https://weekly.blokhin.us
❤️ Patreon: https://www.patreon.com/ZeroToMVP
🖋️ Texodus: https://github.com/w512/Texodus

------------------------------

Previous video, Claude Haiku 5.5 vs local Qwen 3.8 27B on the same task: https://youtu.be/wm2OVdfr_Oo


⏱️ Timestamps
00:00 Intro
00:22 What DeepSeek released
00:54 The task: auto-updating links in Texodus
01:50 DeepSeek V4.1 Flash results
02:20 Claude Opus 5.5 results
02:45 DeepSeek went looking for the answers
04:13 Code quality and edge cases
04:53 Conclusions

#DeepSeek #ClaudeOpus #AICoding</media:description>
   <media:community>
    <media:starRating count="640" average="5.00" min="1" max="5"/>
    <media:statistics views="62006"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:HEnD_UV9UwU</id>
  <yt:videoId>HEnD_UV9UwU</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>I Forced AI to Promote My App (And It Invented Features)</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=HEnD_UV9UwU"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-10-16T15:38:15+00:00</published>
  <updated>2026-10-22T22:02:15+00:00</updated>
  <media:group>
   <media:title>I Forced AI to Promote My App (And It Invented Features)</media:title>
   <media:content url="https://www.youtube.com/v/HEnD_UV9UwU?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i1.ytimg.com/vi/HEnD_UV9UwU/hqdefault.jpg" width="480" height="360"/>
   <media:description>Try Higgsfield Ads Studio: https://higgsfield.ai/s/higgsfield-ads-studio-yt-zerotomvp-rerwnA

I can build apps, but I'm not great at promoting them. So I gave Higgsfield Ads Studio, an AI ad generator, nothing but the website of Kivarion, my free open-source password manager. Static Ads turned it into a brand kit and a batch of static ad creatives, and Ad Multiplier made AI video ads from a screen recording with a new hook and CTA, no reshoot needed. Some ads came out usable, others put my desktop app on a phone or changed its interface, so I checked every one. It's handy for ad creative testing, and not only for ecommerce ads, but like all AI marketing tools, it still needs a human review.

-------------------------------------------------------

📬 Newsletter — https://weekly.blokhin.us
🔐 Kivarion — https://kivarion.blokhin.us

-------------------------------------------------------

⏱️ Chapters:
00:00 I can build apps, but not promote them
01:02 Kivarion and Higgsfield Ads Studio
01:45 Static Ads: from a website to eight ads
03:16 Fixing the brand kit
05:05 Ad Multiplier: video ads without a reshoot
06:27 Conclusions


#AIads #HiggsfieldAI #AdsStudio #AIMarketing #EcommerceMarketing #AdCreative</media:description>
   <media:community>
    <media:starRating count="28" average="5.00" min="1" max="5"/>
    <media:statistics views="26334"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:QS7JCP_c8d8</id>
  <yt:videoId>QS7JCP_c8d8</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>Can a 1B AI Model Actually See? I Tested 29 Tasks</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=QS7JCP_c8d8"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-10-15T13:00:05+00:00</published>
  <updated>2026-10-22T08:20:59+00:00</updated>
  <media:group>
   <media:title>Can a 1B AI Model Actually See? I Tested 29 Tasks</media:title>
   <media:content url="https://www.youtube.com/v/QS7JCP_c8d8?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/QS7JCP_c8d8/hqdefault.jpg" width="480" height="360"/>
   <media:description>Can a 1-billion-parameter vision model actually understand what it sees? I ran MiniCPM-V 4.7 from OpenBMB through 29 real-world tasks (text, objects, charts, reasoning, video, and browser actions) and compared the 1B model with the larger 35B version.

MiniCPM-V 4.7 is small enough to potentially run on a phone or a Raspberry Pi. It read my handwriting, converted a research-paper table to CSV without a single wrong number, followed the sequence of events in a video, and navigated a browser. It also missed a pair of glasses disappearing from a table and called a bright yellow spacebar gray.

Tasks are split into six difficulty levels, and each one runs three times.

-------------------------------------------

🔗 Links
GitHub: http://github.com/OpenBMB/MiniCPM-V 
Hugging Face: http://huggingface.co/collections/openbmb/minicpm-v-47 

-------------------------------------------

⏱️ Timestamps
00:00 Can a 1B model see?
00:30 How the tasks are set up
00:53 Level 1: Reading text (OCR)
01:34 Level 2: Recognizing objects
02:12 Level 3: Charts and diagrams
02:44 Level 4: Reasoning
03:06 Level 5: Video understanding
03:47 Level 6: Taking actions (function calls, browser agent)
04:24 Visual compression up to 16x
04:37 MiniCPM-V 4.7 vs 4.6
04:55 Results and verdict

-------------------------------------------

📬 Newsletter — https://weekly.blokhin.us
❤️ Patreon — https://patreon.com/ZeroToMVP


#MiniCPM #LocalAI #VisionAI</media:description>
   <media:community>
    <media:starRating count="104" average="5.00" min="1" max="5"/>
    <media:statistics views="24127"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:wm2OVdfr_Oo</id>
  <yt:videoId>wm2OVdfr_Oo</yt:videoId>
  <yt:channelId>UCvN3QmpWy_eFG8gIU_Ij-RQ</yt:channelId>
  <title>Claude Haiku 5.5 vs Local Qwen 3.8 27B: Is Local AI Still Worth It?</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=wm2OVdfr_Oo"/>
  <author>
   <name>Zero to MVP</name>
   <uri>https://www.youtube.com/channel/UCvN3QmpWy_eFG8gIU_Ij-RQ</uri>
  </author>
  <published>2026-10-13T16:13:19+00:00</published>
  <updated>2026-10-18T01:09:08+00:00</updated>
  <media:group>
   <media:title>Claude Haiku 5.5 vs Local Qwen 3.8 27B: Is Local AI Still Worth It?</media:title>
   <media:content url="https://www.youtube.com/v/wm2OVdfr_Oo?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i4.ytimg.com/vi/wm2OVdfr_Oo/hqdefault.jpg" width="480" height="360"/>
   <media:description>Claude Haiku 5.5 costs ten cents per million input tokens. So is running a local model on your own GPU still worth it? I gave Haiku 5.5 and a local Qwen 3.8 27B the same four real tasks to find out.

Setup:
- Agent: Pi, same tools for both models, maximum reasoning level
- Cloud: Claude Haiku 5.5 via the Anthropic API
- Local: Qwen 3.8 27B, Q4 quantization, llama.cpp, 192K context
- GPU: AMD Radeon Pro W7800, 32 GB VRAM


📬 Zero to MVP Weekly newsletter: https://weekly.blokhin.us
❤️ Support the channel on Patreon: https://www.patreon.com/ZeroToMVP
🖋️ Texodus: https://github.com/w512/Texodus


Chapters:
00:00 Intro
00:16 Setup
01:15 Task 1: Python expense-splitting utility
02:02 Task 2: Fixing word count in Texodus
02:52 Task 3: Updating links on rename
03:32 Haiku's 100K-token price threshold
05:39 Task 4: Tax deductions from financial documents
06:57 Conclusions

#LocalAI #ClaudeHaiku #Qwen</media:description>
   <media:community>
    <media:starRating count="916" average="5.00" min="1" max="5"/>
    <media:statistics views="89912"/>
   </media:community>
  </media:group>
 </entry>
</feed>
