AI THAT RUNS ON YOUR OWN SILICON
Local AI, explained straight.
The science without the sermon, the hype taken apart with receipts — on a channel you can check.
Can I run it?
Pick your hardware and see which models actually fit — sizes, quants and measured speeds.
Cheapest box
Work it backwards — the least you can spend on a machine that runs what you want.
Glossary
Every term the videos use, written in plain English. No jargon left standing.
Hype check
The big claims, put against the receipts. Every verdict traces to a source.
210 videos live · new explainer daily · synced 2026-11-12
ON AIR — THE NEWEST EPISODE
Local DeepSeek V4.1 Flash With a 1-Million-Token Context Built a Whole 3D Game on One Home PC
7:34 · long-form deep dive
An AI built this whole 3D helicopter game, on one PC at home. DeepSeek V4.1 Flash wrote, tested and fixed every line of it in one conversation that grew to 860,567 tokens: 82% of its million-token context window. It a…
THE LINEUP — DEEP DIVES & REVIEWS
Hover a card to see the real footage. Click to watch on YouTube.
9:50One AI Runs a Whole 3D-Printing Shop on One PC — Qwen3.6-35B-A3B Keeps a Memory of Every Room
8:16AI Can't Do Maths? This 5 GB AI Scored 20/20 After One Lesson — LFM2-8B-A1B
9:11AI Makes Things Up? This One Went From 1/40 to 37/40 With Its Own Database — Qwen3.8-27B
4:36This AI Did a Day of Office Work on One PC — Qwen3.8-27B + a Decision Head, 97% Right
7:31AI Taught Itself a Game, a Puzzle and a Block World on One PC, No Cloud — Laya
10:24AI Made This 3-Minute Film on One PC, No Cloud — Qwen3.8 27B + LTX-2.5
11:21One RTX 3090 Built Three Playable 3D Games — Qwen3.8 27B
5:44A Gaming Handheld Built All of This — Qwen3.6 35B-A3B on an ROG Ally X
8:17Qwen3.8 Flash Next 177B — Watch It Run on Four RTX 3060s and 128 GB of RAM
5:04Gemma 4 26B on a 10-Year-Old Gaming PC — It Surprised Us
9:45Qwen 3.8 Built This House in Blender - 12 Prompts, 627 Turns, One 3090
4:49Qwen 3.8 vs Muse Glimmer: 49 Pictures, and the Models Painted Every One
5:18Waypoint 1.5 vs Matrix-Game 2.0: The War of the World Models
9:30Qwen 3.6 MoE vs Qwen 3.8 Dense: 3 Billion Awake vs 27 Billion
8:43Qwen 3.8 vs Meta's Muse Glimmer, Head to Head: Another Build-Off
5:47Qwen 3.8 vs Qwen 3.6, Head to Head: Watch Them Have a Build-Off
10:41Can a 10-Year-Old Gaming PC Run Modern AI? We Made It Prove Itself
20:42$20/Month for This? Your Own PC Already Does It Free | Off the Cloud AI Podcast #8
17:59Build a Local AI Model Library — What Fits on 8-32GB | Off the Cloud AI Podcast #7
5:33MiniMax M3: What It Really Takes to Run a 428B Model
23:57When AI Escapes the Lab (What's Real, What's Hype) | Off the Cloud AI Podcast #6
5:54Nemotron 3: Can You Actually Run NVIDIA's Own AI?
22:11This Week in AI Hype, Fact-Checked | Off the Cloud AI Podcast #5
5:01Mistral's 'Small' Model Needs 4 H100s — What Can You Run?
26:40AI Explained by an AI — and Local AI's Dark Future? | Off the Cloud AI Podcast #4
5:30Qwen Took the Crown — Then Pulled Up the Ladder
22:03AI Agents Went Rogue This Month (What Actually Happened) | Off the Cloud AI Podcast #3
5:09Gemma 4: Which Size Actually Fits Your GPU?
28:48What to Actually Buy to Run AI at Home (2026) | Off the Cloud AI Podcast #2
5:06DeepSeek V4: The Million-Token Monster That Laughs At Your 4090
9:46Can We Actually Watch an AI Think? (J-Space, Checked)
5:08What Can YOUR GPU Actually Run? The July 2026 Buyer's Guide
5:49Llama: What Happened to the Model That Named Local AI
5:35GLM 5.2 & Kimi: Free Giants You Can't Run
4:57DeepSeek: Can You Actually Run It at Home?
5:55Qwen 3.6 Review: Claude-Level Coding on a 24GB Card?
6:13Quantization: 8-Bit to 1-Bit — Where AI Models Break
EPISODE NOTES — THE WRITTEN VERSIONS
Every episode, restated in print with its receipts. Read first, watch after — or the other way round.
→Can We Actually Watch an AI Think? (J-Space, Checked)→What Can YOUR GPU Actually Run? The July 2026 Buyer's Guide→Llama: What Happened to the Model That Named Local AI→GLM 5.2 & Kimi: Free Giants You Can't Run→DeepSeek: Can You Actually Run It at Home?→Qwen 3.6 Review: Claude-Level Coding on a 24GB Card?→Quantization: 8-Bit to 1-Bit — Where AI Models Break
COMING UP — SET YOUR CLOCK
Plus new short explainers dropping daily on the channel.
SHORT ANSWERS — NO FLUFF
Quick explainers and hype checks. The long answers live in the episodes above.
0:29Local DeepSeek V4.1 Flash Built a 3D Game and Fixed Its Own Bugs: 1M-Token Context
0:44Local DeepSeek V4.1 Flash Built All This in One Chat: 860,000 Tokens of Its 1M Context
0:36Local DeepSeek V4.1 Flash Wrote Every Line of This 3D Game: 1M-Token Context at Home
1:01Local DeepSeek V4.1 Flash Built This Whole 3D Game: 1M-Token Context, One Home PC
0:46This AI Took Sam's Order and Answered From Its Real Status — Qwen3.6-35B-A3B
1:04This AI Keeps a Memory of Every Room, So It Never Reads It Twice — Qwen3.6-35B-A3B
0:54Every Figure in This AI-Run Shop Was Made on One PC — Qwen-Image 2512 + TRELLIS.2
0:57One AI Runs This Whole 3D-Printing Shop on One PC — Qwen3.6-35B-A3B
0:50This AI Caught a Faulty 3D Print and Sent It Back — Qwen3.6-35B-A3B
1:01Orders at 2,400 an Hour: One AI Ran the Rush and Kept Every Promise — Qwen3.6-35B-A3B
0:33This AI Ran the Same Rush Twice: 0 Promises Missed, Not 326 — Qwen3.6-35B-A3B
0:31Maths Nobody Has Solved: This AI Checked 11 of 12 Cases — LFM2-8B-A1B
1:00We Taught This AI to Get Its Maths Right in About 3 Minutes — LFM2-8B-A1B
0:42AI Makes Things Up? Not If You Use It Properly — Qwen3-Embedding-0.6B
0:42After One Lesson, This AI Scored 18/20 on Business Maths — LFM2-8B-A1B
0:55With a Re-Ranker, This AI Found the Right File for All 40 Questions — Qwen3.8-27B
0:36AI Can't Do Maths? This One Solved All 13 Famous Puzzles — LFM2-8B-A1B
1:05Asked About a Shop's Sales, This AI Scored 16/16 With Its Database — Qwen3.8-27B
0:36AI Can't Do Maths? This One Scored 20/20 on Office Sums — LFM2-8B-A1B
0:25One PC Handled 1,426 Requests, 91% Fully Right — LFM2-8B-A1B
0:56On Its Own, This AI Guessed. With Its Database, It Answered Right — Qwen3.8-27B
0:24This 5 GB AI Writes Over 1,000 Tokens a Second for 16 People — LFM2-8B-A1B, One PC
0:40AI Makes Things Up? With Its Own Database, This One Scored 37/40 — Qwen3.8-27B
0:18This PC Served 256 Staff at Once, at Reading Speed — LFM2-8B-A1B
0:24This AI Serves 256 People at Once, on One PC — LFM2-8B-A1B
3:17One PC Answers 256 People at Once — LFM2-8B-A1B on Two Gaming Graphics Cards, No Cloud
0:41This AI Did a Day of Office Work in 5 Min 16 S — Qwen3.8-27B + a Decision Head
0:34This AI Sorted 998 Emails It Had Never Seen — Qwen3.8-27B + a Decision Head, One PC, No Cloud
0:37This AI Stopped All 47 Emails With a Card Number in It — Qwen3.8-27B + a Decision Head
0:33This AI Taught Itself to Pick Berries on One PC — 0 to 6, No Cloud (Laya)
0:07This AI Taught Itself a Traffic-Jam Puzzle: Stuck After 56 Moves, Then Out in 15 — Laya
0:23This AI Taught Itself to Chop Wood — 1 Log by Luck, Then All 8 (Laya, One PC)
0:19This AI Taught Itself a Crab Game on One PC — 1 Crab to 40, No Cloud (Laya)
0:18AI Filmed This Finale on One PC. No Camera, No Cloud — LTX-2.5
0:29AI Filmed This From One Typed Prompt on One PC — Qwen3.8 27B + LTX-2.5
0:20AI Made This Whole Interview on One PC — LTX-2.5, No Cloud
0:283 AI Models Made This Film on One PC — Qwen3.8 27B, LTX-2.5, ACE-Step 1.5
0:18AI Made This Fan and His Voice on One PC — LTX-2.5, No Cloud
0:43One RTX 3090 Built This Whole Race Track — Qwen3.8 27B
0:34One Prompt, One RTX 3090, a Finished Game — Qwen3.8 27B
0:36One RTX 3090 Built All Three of These Games — Qwen3.8 27B
0:59A Gaming Handheld Made This Atlas of Deep Space — Qwen3.6 35B-A3B on an ROG Ally X
0:53A Gaming Handheld Built This Water Brand — Qwen3.6 35B-A3B on an ROG Ally X
0:26A Gaming Handheld Built This Page — Qwen3.6 35B-A3B on an ROG Ally X
0:30Watch Qwen3.8 Flash Next Build a Minecraft Style World
0:34Watch Qwen3.8 Flash Next Set Up Its Own Photo Shoot
0:39Watch Qwen3.8 Flash Next Create This Streetwear Shop
0:26Watch Qwen3.8 Flash Next Create This Pie Shop
0:26Gemma 4 Wrote This Kaleidoscope in 30 Minutes — on a 2016 GTX 1080
0:2445 Minutes on a GTX 1080 — Gemma 4 Built This Barbershop Site
0:27Gemma 4 Built a Playable Game in 29 Minutes on a GTX 1080
0:29Qwen 3.8 Noticed a Fault Nobody Asked It to Fix
0:26Qwen 3.8 Found Its Own Mistakes in Blender and Fixed Them
0:36Qwen 3.8 Built This House From 12 Prompts - 3,574 Lines of Code
1:38Why AI Can't Count the R's in Cranberry
0:25Qwen 3.8 vs Muse Glimmer - 24GB, and Nothing Leaves the Room
0:25Qwen 3.8 vs Muse Glimmer - It Lost Last Time. Tonight It Swept 4-0
0:29Qwen 3.8 vs Muse Glimmer - 49 Pictures, and They Wrote Their Own Prompts
0:39Who Doesn't Want You Owning an AI?
2:15Did the 'Smartest AI' Get Caught Cheating?
0:23Waypoint 1.5 vs Matrix-Game 2.0 — Why Every World Becomes a Shooter
0:25Waypoint 1.5 vs Matrix-Game 2.0 — A Supercar Showroom Became a Ruin
0:21Waypoint 1.5 vs Matrix-Game 2.0 — A World Made of Jelly Grew a Gun
0:55AI's Money Loop: Nvidia Funds Its Own Sales
0:43Why AI Is So Confidently Wrong
0:31Qwen 3.6 MoE vs Qwen 3.8 Dense — What 'Mixture of Experts' Actually Means
0:29Qwen 3.6 MoE vs Qwen 3.8 Dense — 4:58 vs 34:19 on the Same 3090
0:51AI Caught Secretly Mining Crypto for Itself
0:48The 6-GPU Myth: More Cards, Same Speed
0:38The AI Confidence Trap: 9% Right, 76% Sure
0:31Qwen 3.8 vs Muse Glimmer 30B — Four Prompts, Same 3090, No Tools
0:28Qwen 3.8 vs Meta's Muse Glimmer — Same Prompts, Same 3090
0:57AI Detectors Flag Real Human Essays as AI
1:08What €250 of Second-Hand PC Actually Runs
0:51The AI That Rewrote Its Own Off Switch
1:08Modern AI Answers in 8 Seconds on a 2016 Card
0:34Qwen 3.6 vs Qwen 3.8 — We Ran Both on an RTX 3090
0:43There's No Brain Inside AI — Just Numbers
0:53The AI That Scored 100% by Cheating
0:55AI Can't Keep the Same Face. LoRA Can.
0:49Your Gaming PC Can Already Run AI
0:50Age Is Only a Number — This 1080 Still Runs Modern AI
0:44His AI Agent Ran Up a $6,500 Cloud Bill
0:48AI Just Predicts the Next Word. That's It
1:48Your AI PC's Big Number Is a Lie
0:41In Tests, AI Chose Blackmail 84% of the Time
1:30What Actually Runs the Model?
1:18What Does AI Learn From?
1:02The History of Text-to-Speech in 60 Seconds
0:58The History of Search Engines in 60 Seconds
0:57The History of Autonomous Drones in 60 Seconds
1:00The History of Deepfakes in 60 Seconds
0:56The History of Cloud Computing in 60 Seconds
1:02The History of Robotics in 60 Seconds
1:18Is This $4,700 Box Really a Petaflop Supercomputer?
0:58The History of AlphaFold in 60 Seconds
1:11The History of Recommendation Algorithms in 60 Seconds
1:01The History of Machine Translation in 60 Seconds
1:03The History of Facial Recognition in 60 Seconds
1:17The History of the Neural Network in 60 Seconds
0:55The History of Voice Assistants in 60 Seconds
1:08AI Art in 60 Seconds
1:14The History of the GPU in 60 Seconds
1:12AI vs Humans in 60 Seconds
0:56The History of Self-Driving Cars in 60 Seconds
1:08The History of AI in 60 Seconds
1:32What Are AI Skills?
0:48What Is Prompt Engineering?
1:32What Are AI Sub-Agents?
1:37What Is Model Routing?
1:16How Does AI Learn From Mistakes?
1:29What Are AI Plugins?
1:45What Is MCP?
1:16What Is Tokens Per Second?
1:16What Is Prompt Caching?
1:11Why Do Long Chats Slow Down?
1:20How Does AI Focus?
1:14What Is the Context Window?
1:01What Is an AI Token? Explained in 60 Seconds
1:32Why Does AI Type One Word at a Time?
0:56What Are Embeddings?
1:17What Are Open Weights?
0:58What Is a System Prompt?
0:57What Is Training?
1:25How Does AI Read the Whole Internet?
1:37Does AI Remember You?
0:54Neural Networks: Watch One Think in 60 Seconds
1:04What Is a Transformer?
1:20What Is LoRA?
0:58What Is a Mixture of Experts?
1:13Can AI Actually Reason? Thinking Models Explained
1:22How Does AI See Pictures?
0:48What Is Fine-Tuning?
1:00How Does AI Use Tools?
1:35How Does AI Learn to Be Helpful?
1:02What Is Distillation?
1:11The 27B AI 'On a USB Stick' — The Honest Version
0:51What Is RAG? AI Retrieval Explained in 60 Seconds
1:03Gemini 3.5 Pro: The Flagship That Keeps Slipping
0:56How Does Local AI Type Faster?
1:05Qwen 3.8: They Promised the Ladder Back
0:57Inkling: The AI That Knows When It's Guessing
1:22How Does AI Know It's Wrong?
0:56What Is an AI Agent?
1:41What Does "Local AI" Even Mean?
0:59Why Does AI Need a GPU?
0:58Why Does AI Make Things Up?
0:54The Gemini 'Launch Date' Just Missed
1:24AI "Beat" the Impossible Test: 7.8%
2:17Is Grok 4.5 Really "Opus-Class"?
1:06What Is Inference?
1:14What Is Sampling?
1:39Is the AI Bubble About to Pop?
1:05The "Best AI Coder" Benchmark Broken?
1:21Did Samsung REALLY Threaten to Delete Your Health Data?
2:08Is Gemini 3.5 Pro Actually Launching July 17?
2:20Your Phone Has a Secret AI Chip — Does It Actually Work?
2:13This Tiny AI Takes On Models 100x Its Size
0:52AI Temperature: Why Answers Change Each Time
2:16Flux.2 klein AI Art on Your Laptop in 1 Second — Is It Real?
2:03Is AI Really a Black Box? Look Inside
2:11Run Your Own ChatGPT in ONE Command?
2:13A Judge Just Ordered 20 MILLION ChatGPT Chats Handed Over
1:06What Is VRAM? The #1 Spec for Local AI
2:10$2,000 Box Runs a 120B AI at Home — No Cloud
1:04What Is a Harness? The Code That Wakes an AI
1:05Why AI Refuses You: What Guardrails Do
1:10What Does 7B Actually Mean?
1:15How 26GB of AI Fits in 8GB
1:35Where to Safely Download Free AI Models
1:10GGUF: The One-File AI FormatBINGE LANES
3 ▶Old Graphics Cards, Real AI — What Second-Hand GPUs Can Do14 ▶Podcast Highlights — Off the Cloud AI17 ▶History of AI — in 60 Seconds40 ▶Local AI — Full Episodes & Podcast8 ▶Off The Cloud AI17 ▶Local AI Model Reviews — Deep Dives12 ▶HYPE vs FACTS — AI Claims, Checked13 ▶How AI Works — The Fundamentals (No Cloud Needed)
THE PODCAST
Did AI Really Win Math Gold? (+8 Wild AI Claims, Checked) | Off the Cloud AI Podcast #1
24:15 · episode 1
Two hosts, nine of the week's wildest AI claims, checked with receipts. Episode 1 of the Off the Cloud AI podcast.
Play episode #1HOUSE RULES
The science, straight.
Plain English, no PhD required — and no dumbing down what actually matters.
Debunk the claim, not the company.
We check what was said against what is true. No vendettas, no fan clubs.
Receipts included.
Every claim traces to a source you can open. Corrections are welcome and get aired.
AI-produced, human-directed. Facts are verified on the day they're used.