Video wird geladen...

Video konnte nicht geladen werden

Beim Laden dieses Videos ist ein Problem aufgetreten. Dies könnte an einem vorübergehenden Netzwerkproblem liegen oder das Video ist möglicherweise nicht verfügbar.

Video generation now runs right on your laptop. 🎬 Introducing FreeVideo — bringing MiniMax H3 + Video DeltaNet to the hardware you already own. As little as 8GB VRAM + 16GB RAM. ComfyUI integration, LoRAs & custom workflows. Code:

Haocheng Xi 🚗 COLM

2,661 subscribers

182,727 Aufrufe • vor 2 Tagen •via X (Twitter)

Bildung Wissenschaft & Technologie

Anya Rossi• Live Now

Private livecam show

29 Kommentare

Profilbild von Kai Cheung

Kai Cheungvor 2 Tagen

Windows flags the download has virus threat.

Profilbild von Scene Forge

Scene Forgevor 2 Tagen

To create a 1080p video on my RTX 5060 would take about 12 hours 😂

Profilbild von Haocheng Xi 🚗 COLM

Haocheng Xi 🚗 COLMvor 2 Tagen

We sincerely thank Impossible Research for providing computational resources to support the project. Great collaborations with @Andy_ShuoYang @BowenXue_ @JasonFxz @Chenfeng_X 🔥🔥🔥

Profilbild von Qiuyang Mang

Qiuyang Mangvor 2 Tagen

Nice work!

Profilbild von Mr. Decentralized

Mr. Decentralizedvor 2 Tagen

i need this on mac silicon

Profilbild von Vera Ai | Tools & Updates

Vera Ai | Tools & Updatesvor 2 Tagen

Local video generation at 8GB VRAM is impressive progress

Profilbild von Vicks

Vicksvor 2 Tagen

Thank you !!!

Profilbild von Tianrui Feng

Tianrui Fengvor 2 Tagen

Amazing work!

Profilbild von arenosa⚡

arenosa⚡vor 2 Tagen

freevideo.exe is a red flag Do not trust all of the "open source" repos shared in X. Full of scams, threats, danger...

Profilbild von Osiris

Osirisvor 2 Tagen

唇形对齐比画质先过关，口型漂了再好看也留不住。音频驱动再出画，比先出画后硬贴嘴型稳。

Profilbild von Yixin Dong

Yixin Dongvor 2 Tagen

@shiyi_c98 Haocheng ships

Profilbild von Anderson

Andersonvor 2 Tagen

who needs a datacenter for video now

Profilbild von the.krxsna

the.krxsnavor 2 Tagen

seems nice will try it out

Profilbild von Lank

Lankvor 2 Tagen

8GB VRAM MiniMax H3 on the laptop is the number that made me open the repo. ComfyUI + LoRA local before I touch cloud again.

Profilbild von Arafat

Arafatvor 2 Tagen

Running MiniMax H3 locally on 8GB VRAM is incredible work! 🚀 I’d love to feature and promote FreeVideo to my audience. Could you please open your DMs or send me a message so we can discuss a collaboration?

Profilbild von ShadowAguy

ShadowAguyvor 2 Tagen

8GB VRAM and i’m interested.

Profilbild von Magica

Magicavor 2 Tagen

Local inference killing the "rent our GPUs" pitch one repo at a time. Hardware you already own was the unlock nobody priced in.

Profilbild von linlin

linlinvor 2 Tagen

8GB VRAM laptop video generation — my fan spun up just reading this.

Profilbild von YunCuntu | Video & Photo Downloader

YunCuntu | Video & Photo Downloadervor 2 Tagen

8GB VRAM is the interesting part here, local video gen finally reaching consumer cards. Curious how generation time holds up on longer clips.

Profilbild von seedframe

seedframevor 2 Tagen

8GB is tight when the temporal attention cache grows with frame count and resolution. Drop to 24 frames or 480p to stay under budget. Above 32 frames crashes.

Profilbild von Erdal Bektas

Erdal Bektasvor 2 Tagen

@grok does this work at Mac?

Profilbild von X Æ A-12

X Æ A-12vor 2 Tagen

do you verify it's not a trojan program?

Profilbild von MightyLab

MightyLabvor 2 Tagen

@grok 맥북가능?

Profilbild von ningmeng lin

ningmeng linvor 2 Tagen

8GB VRAM for video generation is wild. My laptop fan is already drafting a noise complaint.

Profilbild von Meadows

Meadowsvor 2 Tagen

Will this run on a mac?

Profilbild von Latent Spacer

Latent Spacervor 2 Tagen

Great work guys! You're also great at marketing your stuff!

Profilbild von Shara

Sharavor 2 Tagen

8GB VRAM support with ComfyUI is a big win.

Profilbild von Luis maria

Luis mariavor 2 Tagen

ish

Profilbild von Vini

Vinivor 2 Tagen

Video generation on a laptop is huge.

Ähnliche Videos

introducing the world's first fully ai video streaming service! built using minimax h3 max turbo, you can now watch entire movies and easily create your own ones

introducing the world's first fully ai video streaming service! built using minimax h3 max turbo, you can now watch entire movies and easily create your own ones

Blendi

125,338 Aufrufe • vor 1 Monat

MiniMax H3 is now live on AKOOL! Create cinematic AI videos with synchronized audio, realistic dialogue, and immersive sound effects—all in one generation. Ready for higher-quality AI video creation? Bring your ideas to life with MiniMax H3 on AKOOL today! 🎬

MiniMax H3 is now live on AKOOL! Create cinematic AI videos with synchronized audio, realistic dialogue, and immersive sound effects—all in one generation. Ready for higher-quality AI video creation? Bring your ideas to life with MiniMax H3 on AKOOL today! 🎬

Akool Inc

1,948,818 Aufrufe • vor 2 Monaten

Anyone with 8GB or 12GB VRAM setups needs to understand that "-ncmoe" is the key flag to boost performance on llama.cpp Here are my results for Qwen3.6 35B A3B, with 64k q8_0 context on a 8GB RTX 3070Ti: ⚪️ no flag → 8.7 tok/s RAM: 13.6GB & VRAM: 7.8GB 🔴 -ncmoe 35 → 27.5 tok/s RAM: 12.1GB & VRAM: 4.3GB 🟢 -ncmoe 30 → 32.5 tok/s RAM: 12GB & VRAM: 5.6GB 🔵 -ncmoe 25 → 40.9 tok/s RAM: 12GB & VRAM: 6.9GB Please note the ram and vram usage you see are total usage of a windows pc, with the model running. My friend's setup: 8GB VRAM and 16GB RAM. You can boost performance by switching to Linux, just something to keep in mind. Basically, this flag keeps the MoE experts in the first X layers on your CPU + RAM, instead of eating all your VRAM straight away. This is a smart hybrid offload way that lets you run bigger models without OOM while keeping the rest on your GPU for speed. As we can see on the data, there's a sweet spot. When we lower it from 35 to 25, speed bumps +50% because there are more layers on your GPU (look at the VRAM usage). The key here is to play around with the number and fit as much as possible on your VRAM, goal is to have 1GB/800MB headroom to avoid stress. ↓ server flags below

Anyone with 8GB or 12GB VRAM setups needs to understand that "-ncmoe" is the key flag to boost performance on llama.cpp Here are my results for Qwen3.6 35B A3B, with 64k q8_0 context on a 8GB RTX 3070Ti: ⚪️ no flag → 8.7 tok/s RAM: 13.6GB & VRAM: 7.8GB 🔴 -ncmoe 35 → 27.5 tok/s RAM: 12.1GB & VRAM: 4.3GB 🟢 -ncmoe 30 → 32.5 tok/s RAM: 12GB & VRAM: 5.6GB 🔵 -ncmoe 25 → 40.9 tok/s RAM: 12GB & VRAM: 6.9GB Please note the ram and vram usage you see are total usage of a windows pc, with the model running. My friend's setup: 8GB VRAM and 16GB RAM. You can boost performance by switching to Linux, just something to keep in mind. Basically, this flag keeps the MoE experts in the first X layers on your CPU + RAM, instead of eating all your VRAM straight away. This is a smart hybrid offload way that lets you run bigger models without OOM while keeping the rest on your GPU for speed. As we can see on the data, there's a sweet spot. When we lower it from 35 to 25, speed bumps +50% because there are more layers on your GPU (look at the VRAM usage). The key here is to play around with the number and fit as much as possible on your VRAM, goal is to have 1GB/800MB headroom to avoid stress. ↓ server flags below

left curve dev

166,478 Aufrufe • vor 4 Monaten

🚀 Introducing H3 Reference Turbo Realtime, the latest model by MiniMax (official). Stream video with audio and guide generation with up to 9 reference images, giving you more control than ever. Available today on Reactor. Start building now!

🚀 Introducing H3 Reference Turbo Realtime, the latest model by MiniMax (official). Stream video with audio and guide generation with up to 9 reference images, giving you more control than ever. Available today on Reactor. Start building now!

reactor

2,909,971 Aufrufe • vor 25 Tagen

HERMES AGENT NOW RUNS ON AN 8GB LAPTOP GPU JUST AS EASILY AS IT RUNS ON A 128GB MINI PC Nous Research shipped the official Hermes Agent Desktop App this week. Someone pointed it at a local llama server running on an RTX 4060 with 16GB system RAM. The integration took two minutes The model behind it: Gemma 4 26B MoE, QAT quantized, running on 8GB of VRAM. A 60k token prompt held a stable 20 tokens a second, flat, no slowdown as context grew. The flags were nothing exotic, just -cmoe -c 248000 on llama.cpp What that 8GB setup does out of the box: reads and patches its own code, runs it in a terminal, debugs errors, manages GitHub repos, spawns sub-agents for parallel work. Browses the web with vision to debug a UI. Schedules cron jobs in plain language. Connects to Notion, Google Workspace, Linear, and Obsidian to manage tasks on its own That's the same agent layer running on a Minisforum MS-S1 MAX with 128GB of unified memory, 96GB of it to the GPU, holding a 120B model at 56 tokens a second instead of a 26B model at 20. Same software, same tool execution, same zero API key. The only thing that changes between an $800 laptop and a $2,000 mini PC is how big a model you can afford to run underneath it The barrier to running a real autonomous agent locally didn't just drop. It dropped all the way down to hardware most people already own

HERMES AGENT NOW RUNS ON AN 8GB LAPTOP GPU JUST AS EASILY AS IT RUNS ON A 128GB MINI PC Nous Research shipped the official Hermes Agent Desktop App this week. Someone pointed it at a local llama server running on an RTX 4060 with 16GB system RAM. The integration took two minutes The model behind it: Gemma 4 26B MoE, QAT quantized, running on 8GB of VRAM. A 60k token prompt held a stable 20 tokens a second, flat, no slowdown as context grew. The flags were nothing exotic, just -cmoe -c 248000 on llama.cpp What that 8GB setup does out of the box: reads and patches its own code, runs it in a terminal, debugs errors, manages GitHub repos, spawns sub-agents for parallel work. Browses the web with vision to debug a UI. Schedules cron jobs in plain language. Connects to Notion, Google Workspace, Linear, and Obsidian to manage tasks on its own That's the same agent layer running on a Minisforum MS-S1 MAX with 128GB of unified memory, 96GB of it to the GPU, holding a 120B model at 56 tokens a second instead of a 26B model at 20. Same software, same tool execution, same zero API key. The only thing that changes between an $800 laptop and a $2,000 mini PC is how big a model you can afford to run underneath it The barrier to running a real autonomous agent locally didn't just drop. It dropped all the way down to hardware most people already own

NO1ennn

40,079 Aufrufe • vor 3 Monaten

MiniMax H3 is now on Runware 🎬 - New open-weights flagship video model from MiniMax - Native synced audio, generated in the same pass as the video - Multi-reference conditioning keeps subjects and scene identity consistent across shots - Extend existing clips or audio segments

MiniMax H3 is now on Runware 🎬 - New open-weights flagship video model from MiniMax - Native synced audio, generated in the same pass as the video - Multi-reference conditioning keeps subjects and scene identity consistent across shots - Extend existing clips or audio segments

Runware

10,913 Aufrufe • vor 2 Monaten

MiniMax H3 is native in ComfyUI, day zero with the weights. Model highlights: → Text-to-video, prompt only → Image-to-video → First-and-last-frame - control the opening frame, the closing frame, or both → Reference-to-video - carry a subject, a motion, or a voice through the clip → In-place editing - modify a shot you already have instead of regenerating it The capability MiniMax (official) leads with is multimodal context understanding, and it's what collapses those five tasks into one model. H3 takes images, audio, and video together and resolves them against a prompt that explains how they relate. Download the workflows & learn more in our blog 👇

MiniMax H3 is native in ComfyUI, day zero with the weights. Model highlights: → Text-to-video, prompt only → Image-to-video → First-and-last-frame - control the opening frame, the closing frame, or both → Reference-to-video - carry a subject, a motion, or a voice through the clip → In-place editing - modify a shot you already have instead of regenerating it The capability MiniMax (official) leads with is multimodal context understanding, and it's what collapses those five tasks into one model. H3 takes images, audio, and video together and resolves them against a prompt that explains how they relate. Download the workflows & learn more in our blog 👇

ComfyUI

235,022 Aufrufe • vor 2 Monaten

Comfy MCP is now local and open-source! The #1 ask after Comfy Cloud MCP shipped in June. Your agent sees your actual install. So every node, custom nodes included, and every model on disk. It fetches files, starts your instance, gets a workflow to the point where it runs. It is now the easiest way to help with your local Minimax H3 workflows! Cloud MCP still does everything it did. Tell your agent where a job goes, or let it decide.

Comfy MCP is now local and open-source! The #1 ask after Comfy Cloud MCP shipped in June. Your agent sees your actual install. So every node, custom nodes included, and every model on disk. It fetches files, starts your instance, gets a workflow to the point where it runs. It is now the easiest way to help with your local Minimax H3 workflows! Cloud MCP still does everything it did. Tell your agent where a job goes, or let it decide.

ComfyUI

184,806 Aufrufe • vor 1 Monat

MiniMax H3 is the SOTA open weight model... but the large model comes with slow gens (depending on your hardware). You can be waiting for minutes just for the generation to miss the mark Here's a ComfyUI workflow to get the most out of H3 without the wait. → H3 Turbo LoRA (4–8 steps) → Model Preview Override + Tiny AutoEncoder H3 for a fast composition preview → Bad preview? stop it, tweak the prompt or seed → Save your precious gpu time Workflow and model links below 👇

MiniMax H3 is the SOTA open weight model... but the large model comes with slow gens (depending on your hardware). You can be waiting for minutes just for the generation to miss the mark Here's a ComfyUI workflow to get the most out of H3 without the wait. → H3 Turbo LoRA (4–8 steps) → Model Preview Override + Tiny AutoEncoder H3 for a fast composition preview → Bad preview? stop it, tweak the prompt or seed → Save your precious gpu time Workflow and model links below 👇

rob - comfyui

26,990 Aufrufe • vor 1 Monat

$a regular laptop with just 8GB VRAM can now run a LOCAL AI AGENT directly on your device what you need: 1/ 8 GB VRAM 2/ 16 GB RAM 3/ $600+ laptop method: Kimi K2 + Hermes Agent Dekstop App what agent can do: 1/ read and write files 2/ patch code, open PRs on GitHub 3/ browse the web like a human 4/ work with Notion 5/ run scripts, debug, refactor this isn't 'ask a question, get an answer.' this is an agent that drives a task from start to finish on its own. before: local agent = either the cloud and $$$, or an RTX now: any gaming laptop from the last 3 years p.s. MoE is the key. Kimi K2 activates only a fraction of its params per token - that's why a 1T-param model runs on hardware where a dense 26B would already be choking did you try Hermes Dekstop?$

a regular laptop with just 8GB VRAM can now run a LOCAL AI AGENT directly on your device what you need: 1/ 8 GB VRAM 2/ 16 GB RAM 3/ $600+ laptop method: Kimi K2 + Hermes Agent Dekstop App what agent can do: 1/ read and write files 2/ patch code, open PRs on GitHub 3/ browse the web like a human 4/ work with Notion 5/ run scripts, debug, refactor this isn't 'ask a question, get an answer.' this is an agent that drives a task from start to finish on its own. before: local agent = either the cloud and $$$, or an RTX now: any gaming laptop from the last 3 years p.s. MoE is the key. Kimi K2 activates only a fraction of its params per token - that's why a 1T-param model runs on hardware where a dense 26B would already be choking did you try Hermes Dekstop?

kaize

58,760 Aufrufe • vor 3 Monaten

Open-source video generation is now faster than playback without compromising quality. Introducing Video Delta Net (VDN): hybrid attention for live text-to-video with near-lossless quality. VDN accelerates Minimax-H3 by 75 - 90 x, generating 14 seconds of 768p video in 11 seconds on 8× NVIDIA B200 GPUs. Checkpoints + training/inference code + Technical Blog ⬇️ (1/6)

Open-source video generation is now faster than playback without compromising quality. Introducing Video Delta Net (VDN): hybrid attention for live text-to-video with near-lossless quality. VDN accelerates Minimax-H3 by 75 - 90 x, generating 14 seconds of 768p video in 11 seconds on 8× NVIDIA B200 GPUs. Checkpoints + training/inference code + Technical Blog ⬇️ (1/6)

Haocheng Xi 🚗 COLM

442,010 Aufrufe • vor 1 Monat

Testing LCM LORAs in an AnimateDiff & multi-controlnet workflow in ComfyUI. I was able to process this entire Black Pink music video as a single .mp4 input. The LCM lets me render at 6 steps (vs 20+) on my 4090 and uses up only 10.5 GB of VRAM. Here's a breakdown 🧵[1/11]

Testing LCM LORAs in an AnimateDiff & multi-controlnet workflow in ComfyUI. I was able to process this entire Black Pink music video as a single .mp4 input. The LCM lets me render at 6 steps (vs 20+) on my 4090 and uses up only 10.5 GB of VRAM. Here's a breakdown 🧵[1/11]

CoffeeVectors

182,476 Aufrufe • vor 2 Jahren

I know some bros want to see what happens when we push the value even further, so here we go: ⚪️ -ncmoe 25 → 41.8 tok/s RAM: 12GB & VRAM: 6.9GB 🔴 -ncmoe 23 → 43.8 tok/s RAM: 12.2GB & VRAM: 7.4GB 🟢 -ncmoe 21 → 38.6 tok/s RAM: 12.4GB & VRAM: 7.8GB 🔵 -ncmoe 19 → 19.8 tok/s RAM: 13.8GB & VRAM: 7.8GB As you can see, there's a sweet spot with the VRAM usage. Play around and monitor to find the right value for your setup, you can use llama.cpp web ui to monitor speeds easily Sweet spot seems to be 25-23 for 8GB VRAM ✅ +40tok/s for Qwen3.6 35B with 64k q8_0 context on a 8GB card is very impressive just by using base llama.cpp, and we didn't even try Turboquant, MTP or Dflash yet! I'll focus on these next 👀 (Server flags and setup in the quoted tweet)

I know some bros want to see what happens when we push the value even further, so here we go: ⚪️ -ncmoe 25 → 41.8 tok/s RAM: 12GB & VRAM: 6.9GB 🔴 -ncmoe 23 → 43.8 tok/s RAM: 12.2GB & VRAM: 7.4GB 🟢 -ncmoe 21 → 38.6 tok/s RAM: 12.4GB & VRAM: 7.8GB 🔵 -ncmoe 19 → 19.8 tok/s RAM: 13.8GB & VRAM: 7.8GB As you can see, there's a sweet spot with the VRAM usage. Play around and monitor to find the right value for your setup, you can use llama.cpp web ui to monitor speeds easily Sweet spot seems to be 25-23 for 8GB VRAM ✅ +40tok/s for Qwen3.6 35B with 64k q8_0 context on a 8GB card is very impressive just by using base llama.cpp, and we didn't even try Turboquant, MTP or Dflash yet! I'll focus on these next 👀 (Server flags and setup in the quoted tweet)

left curve dev

18,049 Aufrufe • vor 4 Monaten

Introducing ComfyUI in Unity demo - Supercharges your game development with open-source gen AI. Run any creative workflow in the cloud, and integrative nicely to your Unity scene view port. Imaging what you can already do with ComfyUI - texture, sprite generation - 3d generation - llm, story generation You can now scale, and build tools around these creative workflows. And integrate it into any domain-specific pipeline. Inspired by Magnific.ai ! Running serverless with Modal p.s. I'm still in GDC, if anyone is interested please reach out. Will be testing with more workflows this week.

Introducing ComfyUI in Unity demo - Supercharges your game development with open-source gen AI. Run any creative workflow in the cloud, and integrative nicely to your Unity scene view port. Imaging what you can already do with ComfyUI - texture, sprite generation - 3d generation - llm, story generation You can now scale, and build tools around these creative workflows. And integrate it into any domain-specific pipeline. Inspired by Magnific.ai ! Running serverless with Modal p.s. I'm still in GDC, if anyone is interested please reach out. Will be testing with more workflows this week.

BennyKok

57,970 Aufrufe • vor 2 Jahren

Haha! This is crazy. Maestro v1.8.5 w/ MiniMax-Music3 just dropped. As a test, I prompted "Wheels on the Bus as a death metal song by the band Nursery Crimes." But it's Maestro, so I told it to auto-make the entire video using MiniMax-H3 for video.🤯 Free, Local & Open! 🔊🎵📽️

Haha! This is crazy. Maestro v1.8.5 w/ MiniMax-Music3 just dropped. As a test, I prompted "Wheels on the Bus as a death metal song by the band Nursery Crimes." But it's Maestro, so I told it to auto-make the entire video using MiniMax-H3 for video.🤯 Free, Local & Open! 🔊🎵📽️

Blaine Brown

376,188 Aufrufe • vor 1 Monat

#Viggle is back with an open model — Viggle-Animate — and it’s pretty damn capable. For the first time, they’re releasing weights + code. It replaces characters in video using MiniMax H3. Take a video, grab one frame, manually redraw the character as anything you want, and the model propagates that replacement through the entire clip while preserving motion, camera, and timing. No pose skeleton, segmentation, face tracking, depth, or text prompt. Just the video + one modified frame. Under the hood: a 33.1B MiniMax H3 ref2va fine-tune + DMD2 LoRA. Distillation brings it down to 4 sampling steps. It’s 6× faster than Wan2.2-Animate-14B. The catch: hardware. BF16 needs ~62 GiB for the transformer and peaks at ~80.1 GiB VRAM. They recommend 96 GB+ or CPU offload. It reportedly runs on a 32 GB RTX 5090 via NVFP4, but that quant isn’t public yet. Also, you can’t use a normal character portrait. The reference has to be a frame from the original video, already redrawn as the target character. #ai #web3

#Viggle is back with an open model — Viggle-Animate — and it’s pretty damn capable. For the first time, they’re releasing weights + code. It replaces characters in video using MiniMax H3. Take a video, grab one frame, manually redraw the character as anything you want, and the model propagates that replacement through the entire clip while preserving motion, camera, and timing. No pose skeleton, segmentation, face tracking, depth, or text prompt. Just the video + one modified frame. Under the hood: a 33.1B MiniMax H3 ref2va fine-tune + DMD2 LoRA. Distillation brings it down to 4 sampling steps. It’s 6× faster than Wan2.2-Animate-14B. The catch: hardware. BF16 needs ~62 GiB for the transformer and peaks at ~80.1 GiB VRAM. They recommend 96 GB+ or CPU offload. It reportedly runs on a 32 GB RTX 5090 via NVFP4, but that quant isn’t public yet. Also, you can’t use a normal character portrait. The reference has to be a frame from the original video, already redrawn as the target character. #ai #web3

VORTEX: AI Bros & AI Arena, Peak AI Buzz

47,030 Aufrufe • vor 27 Tagen

For years, local AI had one answer: buy an NVIDIA GPU. RTX 3090. 24GB VRAM. RTX 4090. 24GB VRAM. RTX 5090. 32GB VRAM. And if you wanted to run something bigger, the price climbed fast. Then this video shows a different reality. An old M1 MacBook Air. 16GB unified memory. A DeepSeek R1 0528 Qwen3 8B model. Just 4.62GB on disk. Then the context gets pushed to almost 50,000 tokens. And the 16GB machine stays in the green. Meanwhile, the 8GB Mac is already sitting at 7.51GB used out of 8GB. That's the part I think people still underestimate. Local AI isn't one giant hardware market anymore. A $4,000 RTX 5090 setup, a 96GB Strix Halo box and an old M1 MacBook aren't competing for the same job. They're different rungs. The mistake is buying hardware for the model you wish you needed instead of the work you actually do. If an 8B model can handle your coding, documents, automation and daily tasks, why are you pricing a 32GB GPU? And that's where the subscription math gets interesting. A lot of people are still paying every month for AI without ever checking what the laptop already sitting on their desk can do. The best local AI machine might not be the fastest one. It might be the one you already own.

For years, local AI had one answer: buy an NVIDIA GPU. RTX 3090. 24GB VRAM. RTX 4090. 24GB VRAM. RTX 5090. 32GB VRAM. And if you wanted to run something bigger, the price climbed fast. Then this video shows a different reality. An old M1 MacBook Air. 16GB unified memory. A DeepSeek R1 0528 Qwen3 8B model. Just 4.62GB on disk. Then the context gets pushed to almost 50,000 tokens. And the 16GB machine stays in the green. Meanwhile, the 8GB Mac is already sitting at 7.51GB used out of 8GB. That's the part I think people still underestimate. Local AI isn't one giant hardware market anymore. A $4,000 RTX 5090 setup, a 96GB Strix Halo box and an old M1 MacBook aren't competing for the same job. They're different rungs. The mistake is buying hardware for the model you wish you needed instead of the work you actually do. If an 8B model can handle your coding, documents, automation and daily tasks, why are you pricing a 32GB GPU? And that's where the subscription math gets interesting. A lot of people are still paying every month for AI without ever checking what the laptop already sitting on their desk can do. The best local AI machine might not be the fastest one. It might be the one you already own.

KyzoroX

67,357 Aufrufe • vor 1 Monat

The next generation of HyperFrames skills is here 9 workflows your agent now actually understands a launch video. a music video. captions and overlays on video it knows which one you mean from context and routes there on its own get started below ↓

The next generation of HyperFrames skills is here 9 workflows your agent now actually understands a launch video. a music video. captions and overlays on video it knows which one you mean from context and routes there on its own get started below ↓

HeyGen

1,593,302 Aufrufe • vor 3 Monaten

This video was almost faster to make than to watch Unlimited MiniMax H3 Max for the next 3 days, exclusively in Magnific → Generate as much as you want at 480p → Pay credits to upgrade to 768p and 2K at full speed → For Premium+ and Pro users Available now on Magnific

This video was almost faster to make than to watch Unlimited MiniMax H3 Max for the next 3 days, exclusively in Magnific → Generate as much as you want at 480p → Pay credits to upgrade to 768p and 2K at full speed → For Premium+ and Pro users Available now on Magnific

Magnific

71,072 Aufrufe • vor 1 Monat

a new 8GB VRAM GPU dense Local LLM leader was born yesterday runs on: RTX 4060 / RTX 3070 / RTX 2080. any 8GB card Qwen 3.5 9B (dense) was the go to for 6-8GB VRAM builds. Gemma 4 12B QAT (dense) just changed that. same llama.cpp + cuda 13.2. i7 12700H. 16GB RAM. same -ngl 99 flags. same 48k context. unsloth gemma-4-12b-it-Q4_K_M.gguf → 15 tok/sec @ 48k ctx unsloth gemma-4-12B-it-qat-UD-Q4_K_XL.gguf → 32 tok/sec @ 48k ctx → 26 tok/sec @ 64k ctx 64k context is a big deal. Hermes 3 agent requires 64k minimum to run. you're now getting full hermes compatible context on a budget consumer GPU at 26 tok/sec locally. 2.1x faster on identical hardware. and here's the part that breaks your brain: the QAT-UD-Q4_K_XL is actually SMALLER than the Q4_K_M "XL" why? QAT = Quantization Aware Training Google didn't train the model first and compress it later they trained it to be quantized from day one the weights already know how to survive low precision that's why you get more quality per byte llamacpp flags: -m gemma-4-12B-it-qat-UD-Q4_K_XL.gguf -cnv -ngl 99 -c 48000 -v fits in 8GB VRAM clean. no API. no cloud. no subscription. and this isn't even the MTP variant yet Gemma-4-E2B QAT runs on 3GB RAM, E4B on 5GB, 12B on 7GB, 26-A4B on 15GB and 31B on 18GB. I have benchmarked the 26b and 31b qat as well on a single RTX 4090, checkout the comments for details. If you have a 6GB or 8GB VRAM GPU, post your numbers. more benchmarks and configs coming soon

a new 8GB VRAM GPU dense Local LLM leader was born yesterday runs on: RTX 4060 / RTX 3070 / RTX 2080. any 8GB card Qwen 3.5 9B (dense) was the go to for 6-8GB VRAM builds. Gemma 4 12B QAT (dense) just changed that. same llama.cpp + cuda 13.2. i7 12700H. 16GB RAM. same -ngl 99 flags. same 48k context. unsloth gemma-4-12b-it-Q4_K_M.gguf → 15 tok/sec @ 48k ctx unsloth gemma-4-12B-it-qat-UD-Q4_K_XL.gguf → 32 tok/sec @ 48k ctx → 26 tok/sec @ 64k ctx 64k context is a big deal. Hermes 3 agent requires 64k minimum to run. you're now getting full hermes compatible context on a budget consumer GPU at 26 tok/sec locally. 2.1x faster on identical hardware. and here's the part that breaks your brain: the QAT-UD-Q4_K_XL is actually SMALLER than the Q4_K_M "XL" why? QAT = Quantization Aware Training Google didn't train the model first and compress it later they trained it to be quantized from day one the weights already know how to survive low precision that's why you get more quality per byte llamacpp flags: -m gemma-4-12B-it-qat-UD-Q4_K_XL.gguf -cnv -ngl 99 -c 48000 -v fits in 8GB VRAM clean. no API. no cloud. no subscription. and this isn't even the MTP variant yet Gemma-4-E2B QAT runs on 3GB RAM, E4B on 5GB, 12B on 7GB, 26-A4B on 15GB and 31B on 18GB. I have benchmarked the 26b and 31b qat as well on a single RTX 4090, checkout the comments for details. If you have a 6GB or 8GB VRAM GPU, post your numbers. more benchmarks and configs coming soon

Alok

264,062 Aufrufe • vor 4 Monaten