Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

I made Blender node previews work on ANY PC… No GPU? Still fast. Low-end setup? Still smooth. High-end GPU? Even better. Blender 4.0+ supported #b3d #blender3d #gamedev #3dart

Mix 3D

2,816 subscribers

30,823 views • 5 months ago •via X (Twitter)

Arts Science & Technology #b3d #blender3d #gamedev #3dart

Anya Rossi• Live Now

Private livecam show

26 Comments

Mix 3D5 months ago

its available now with early birds discount :

Mix 3D5 months ago

the demo shows low res there is more resolutions to choose from

Precious Cypher 🥇5 months ago

So when I assigned materials to a model I see the previews ?

Mix 3D5 months ago

you can actually view the last node for the final result

Precious Cypher 🥇5 months ago

Hmm interesting Does it take the shape of the model ?

Mix 3D5 months ago

its a shader based preview it actually nap it to a plane to view it fast so you can see what the mixture of nodes your making as quick as possible while you're on full screen for example

VersaTale5 months ago

This is super fast and responsive compared to Node Preview(some other addon I have)

Mix 3D5 months ago

this one is gpu based the other addons i saw is cpu based which is slow

James George5 months ago

That's awesome 😎

Mix 3D5 months ago

oh thanks James

Precious Cypher 🥇5 months ago

How can I get it ?

Mix 3D5 months ago

it's going to be available in next hours at

Caleb_trynabeanartist5 months ago

Extremely cool.

Mix 3D5 months ago

thanks i will share it when it's done on my page

Thomas Schmall5 months ago

Super neat - it works in blender 5.1 in first little tryouts I did. I wasn't quite sure which of the two files is the newest version. Maybe that could be made more clear. Also it says it needs cycles, but it did work with Eevee so far.

Mix 3D5 months ago

i hope you liked it and if you face any issues just type it down in the post comments so i fix it.

Thomas Schmall5 months ago

Thanks. I will. It showed only black previews at the beginning. A "clear cache" fixed it - since then it's going strong 🙂

Mix 3D5 months ago

thanks mate for feedback.

QUI5 months ago

looks cool

Mix 3D5 months ago

thank you Qui

Serac1115 months ago

Works with GeoNodes?

Mix 3D5 months ago

it only works for shaders. GeoNodes are dealing with meshes and stuff like this it's better to not use something like this for it

Rhal5 months ago

This unbelievably complex tech will be officially added in 10 years.

Mix 3D5 months ago

the complex part is actually getting it to be that fast and crisp looking the dev of this addon started in 2025 the first one was slow as heck and not even good. and now look at that works like a champ

Rhal5 months ago

Amazing work. Hopefully this will work with any node.

Mix 3D5 months ago

haven't tested all of them but i can say works with all nodes i have used in my journey with blender so should be fine. if there is any probem with it just send a report to my email so i can fix it. here will i post it

Related Videos

Guys, Maybe I'm mistaken, but I think there is no way to disable a Light on Blender. That's why I made a toggle button for my Better Pie Menus Add-on. Please tell me what you think. #b3d #blender3d #blender

Guys, Maybe I'm mistaken, but I think there is no way to disable a Light on Blender. That's why I made a toggle button for my Better Pie Menus Add-on. Please tell me what you think. #b3d #blender3d #blender

Pitiwazou 3D Freelancer

11,753 views • 1 year ago

Introducing Kitten TTS, a SOTA tiny text-to-speech model - Just 15M parameters - Runs without a GPU - Model size less than 25 MB - Multiple high-quality voices - Ultra-fast - even runs on low-end edge devices Github and HF links below

Introducing Kitten TTS, a SOTA tiny text-to-speech model - Just 15M parameters - Runs without a GPU - Model size less than 25 MB - Multiple high-quality voices - Ultra-fast - even runs on low-end edge devices Github and HF links below

Divam Gupta

348,944 views • 1 year ago

I made a RuneScape-like game for Mac-in full 3D-using GPT-6 Astra, the Tripo plugin & Blender! This is only version 0.1, so there’s still a lot to improve, but the full 3D animations already work, the core game actions are functional & there’s even a flying dragon at the end!

I made a RuneScape-like game for Mac-in full 3D-using GPT-6 Astra, the Tripo plugin & Blender! This is only version 0.1, so there’s still a lot to improve, but the full 3D animations already work, the core game actions are functional & there’s even a flying dragon at the end!

Derya Unutmaz, MD

21,741 views • 3 days ago

Here's something I promised to show you… GPU grass that is actually fast and usable! It can be dynamically scattered over any mesh in your scene, can be bent by people's avatars or props, and it even works on Quest and mobile just fine. Still WIP, but you can check this world:

Here's something I promised to show you… GPU grass that is actually fast and usable! It can be dynamically scattered over any mesh in your scene, can be bent by people's avatars or props, and it even works on Quest and mobile just fine. Still WIP, but you can check this world:

RED_SIM

47,961 views • 8 months ago

Marvel Tokon 8/10 PC Update Win11 (NYCFurby Framework fgcOS 11) CPU: i9-7980xe GPU: RTX3090 Task Manager Open Youtube: Overall: Significant Upgrade over Launch version. Pachi's team moved fast here. Unsurprising. Issues: Still fairly high overhead and Process Usage (7 PSN services, anticheat etc.) There is probably room for optimization here similar to DBFZ on PC (s/o to Pokevirus) Torwards the end of the video: WHAT IS NGX UPDATER? Task Manager open... there are some EAC Anticheat issues. Tokon has SIGNIFICANT processes updating before game starts. Overall: significant update. Bonus: I have frame generation on for 120fps capture. The 1 netcode match I had was solid. PC Marvel Tokon (with frame generation) > PS5 You're welcome, FGC

Marvel Tokon 8/10 PC Update Win11 (NYCFurby Framework fgcOS 11) CPU: i9-7980xe GPU: RTX3090 Task Manager Open Youtube: Overall: Significant Upgrade over Launch version. Pachi's team moved fast here. Unsurprising. Issues: Still fairly high overhead and Process Usage (7 PSN services, anticheat etc.) There is probably room for optimization here similar to DBFZ on PC (s/o to Pokevirus) Torwards the end of the video: WHAT IS NGX UPDATER? Task Manager open... there are some EAC Anticheat issues. Tokon has SIGNIFICANT processes updating before game starts. Overall: significant update. Bonus: I have frame generation on for 120fps capture. The 1 netcode match I had was solid. PC Marvel Tokon (with frame generation) > PS5 You're welcome, FGC

NYCFurby Framework (Phase II FGC v1.01a)

15,500 views • 1 month ago

Assassins Creed: Shadows PC Port is absolute garbage. I'm playing AC: Shadows both on Console and PC, unfortunately even if you have a high-end PC, the Game stutters on PC. This is most likely because of Denuvo and bad Optimization. The Frametime is so bad I can't even begin to describe it, I'm forced to turn Frame-Gen on just to have a smooth experience. Disabling Frame-Gen and even running this Game on the lowest possible settings still results in a sub-par experience. This clip credit to zWORMz Gaming shows him trying to run AC: Shadows with Max Settings on an RTX 5070 Ti with the Ryzen 7 9800X3D and the Game still struggles to hit 60FPS! having specs anything even below these ones means that the Game will run even worse on most PC Gamers. The PC Port is terrible.

Assassins Creed: Shadows PC Port is absolute garbage. I'm playing AC: Shadows both on Console and PC, unfortunately even if you have a high-end PC, the Game stutters on PC. This is most likely because of Denuvo and bad Optimization. The Frametime is so bad I can't even begin to describe it, I'm forced to turn Frame-Gen on just to have a smooth experience. Disabling Frame-Gen and even running this Game on the lowest possible settings still results in a sub-par experience. This clip credit to zWORMz Gaming shows him trying to run AC: Shadows with Max Settings on an RTX 5070 Ti with the Ryzen 7 9800X3D and the Game still struggles to hit 60FPS! having specs anything even below these ones means that the Game will run even worse on most PC Gamers. The PC Port is terrible.

NikTek

752,324 views • 1 year ago

2025 was a bit of a special year for me. lots of change, lots of personality-defining moment, and lots of growth as well. i want to thank all my friends: work friends, party friends, friend friends, and especially friends in the intersection of these friend categories (you know who you are) <3 i would like to thank all partners, clients, and everyone else i might have come across in 2025. going into 2026, i'm hopeful that we will be even better and continue to improve ourselves. big things coming! <- i am serious anyway, i made this scene quickly to take a little break from work :) i designed this character with Dimas at the end of 2025, and it looks like i might see her more often in my personal works throughout 2026 <3 created in Blender, Houdini, and DaVinci Resolve. The usual suspects. cheers! -i #b3d #blender3d #houdini #sidefxhoudini

2025 was a bit of a special year for me. lots of change, lots of personality-defining moment, and lots of growth as well. i want to thank all my friends: work friends, party friends, friend friends, and especially friends in the intersection of these friend categories (you know who you are) <3 i would like to thank all partners, clients, and everyone else i might have come across in 2025. going into 2026, i'm hopeful that we will be even better and continue to improve ourselves. big things coming! <- i am serious anyway, i made this scene quickly to take a little break from work :) i designed this character with Dimas at the end of 2025, and it looks like i might see her more often in my personal works throughout 2026 <3 created in Blender, Houdini, and DaVinci Resolve. The usual suspects. cheers! -i #b3d #blender3d #houdini #sidefxhoudini

izzy

13,817 views • 8 months ago

just built a local AI waifu on my two macbooks by using the oss parallax inference library!!! not super smooth yet but parallax is still the easiest way to run llms locally on top of mlx - and no bigco can steal my waifu. still need folks to help her see/hear/draw since mlx has no legit wrapper yet. i’m running a campaign giving away mac minis + dgx spark. please please share your localhost ai model setup by using parallax (you don't need a gpu to try self-hosting you can just start with an extra macbook you borrowed from your friend like i did) or any fun app built on top of yourself host model to link below. come build sovereign intelligence with me 🫶

just built a local AI waifu on my two macbooks by using the oss parallax inference library!!! not super smooth yet but parallax is still the easiest way to run llms locally on top of mlx - and no bigco can steal my waifu. still need folks to help her see/hear/draw since mlx has no legit wrapper yet. i’m running a campaign giving away mac minis + dgx spark. please please share your localhost ai model setup by using parallax (you don't need a gpu to try self-hosting you can just start with an extra macbook you borrowed from your friend like i did) or any fun app built on top of yourself host model to link below. come build sovereign intelligence with me 🫶

Asuka Zheng🎀@ICML

178,998 views • 9 months ago

Today we're taking a very early look at the progress being made on the Sega Dreamcast's very own port of Runescape Classic! If you weren't already aware of the work that developer, Lobotomy, has been doing, here's some footage captured from an emulator of some online play! It's still extremely early, and as with every ambitious port in the beginning, it still requires the 32MB RAM upgrade, has no audio, and runs like ass... So basically standard procedure! He's currently using our SDL2 back-end with software rendering, which explains the low framerate. Eventually it will move to HW accelerated graphics using the DC's PowerVR GPU, so the current low FPS is to be expected until that time. What IS a big deal is that even in this early stage, it's already able to connect to the online servers for cross-play and is fully playable with the mouse and keyboard! Stay tuned for future updates as Lobotomy continues his work!

Today we're taking a very early look at the progress being made on the Sega Dreamcast's very own port of Runescape Classic! If you weren't already aware of the work that developer, Lobotomy, has been doing, here's some footage captured from an emulator of some online play! It's still extremely early, and as with every ambitious port in the beginning, it still requires the 32MB RAM upgrade, has no audio, and runs like ass... So basically standard procedure! He's currently using our SDL2 back-end with software rendering, which explains the low framerate. Eventually it will move to HW accelerated graphics using the DC's PowerVR GPU, so the current low FPS is to be expected until that time. What IS a big deal is that even in this early stage, it's already able to connect to the online servers for cross-play and is fully playable with the mouse and keyboard! Stay tuned for future updates as Lobotomy continues his work!

Falco Girgis

11,835 views • 1 year ago

Unity shipped NEW official plugin for AI agents. They released it for Claude Code, but of course I tested it with Codex CLI. GPT-6 is better now😁 Setup got quicker: it's on the Codex marketplace too, one command to hook up. You still install the package with the CLI inside Unity, the simplification is on the agent side. I've been on a custom Unity MCP for months, so I wanted to see if the official become better. It did improved. My test setup: Unity MCP + Blender MCP together. The agent generated models, rigged and animated them in Blender, then wired everything into Unity. For rigs and animations this is flat out the best setup I've tried: just give it Blender via MCP! Watching GPT-6 autotest the game is something else. It writes autotests fast, they catch real bugs, and on my recording it's blasting through test runs like a maniac. 27 skills right now, and the Unity team says more updates are coming. If you want to build your game with AI today, pick Unity. Everything is simply faster. Unreal works too, but it's a much harder road. Source - Install:

Unity shipped NEW official plugin for AI agents. They released it for Claude Code, but of course I tested it with Codex CLI. GPT-6 is better now😁 Setup got quicker: it's on the Codex marketplace too, one command to hook up. You still install the package with the CLI inside Unity, the simplification is on the agent side. I've been on a custom Unity MCP for months, so I wanted to see if the official become better. It did improved. My test setup: Unity MCP + Blender MCP together. The agent generated models, rigged and animated them in Blender, then wired everything into Unity. For rigs and animations this is flat out the best setup I've tried: just give it Blender via MCP! Watching GPT-6 autotest the game is something else. It writes autotests fast, they catch real bugs, and on my recording it's blasting through test runs like a maniac. 27 skills right now, and the Unity team says more updates are coming. If you want to build your game with AI today, pick Unity. Everything is simply faster. Unreal works too, but it's a much harder road. Source - Install:

Stefan 3D AI

41,896 views • 2 days ago

I got Llama 3 running in my browser using only my GPU with my Wi-Fi switched OFF completely client-side WebGPU is a new feature in browsers where JS can use the GPU of the device and apparently you can run LLMs on it too, and it's fast! My end goal for 🧠 is a locally run client side 100% private conversation where nothing gets sent to, processed in or stored in the cloud This will be helpful because I prefer making it web based instead of coding for 4 different native platforms (iOS, Android, MacOS, Windows) Two challenges still: 1) the user has to download the LLM model first and the smallest model is still ~3GB, but they only have to download it once 2) not all devices have fast GPUs yet for LLMs, but it's starting to become a default in most devices e.g.s smartphones

I got Llama 3 running in my browser using only my GPU with my Wi-Fi switched OFF completely client-side WebGPU is a new feature in browsers where JS can use the GPU of the device and apparently you can run LLMs on it too, and it's fast! My end goal for 🧠 is a locally run client side 100% private conversation where nothing gets sent to, processed in or stored in the cloud This will be helpful because I prefer making it web based instead of coding for 4 different native platforms (iOS, Android, MacOS, Windows) Two challenges still: 1) the user has to download the LLM model first and the smallest model is still ~3GB, but they only have to download it once 2) not all devices have fast GPUs yet for LLMs, but it's starting to become a default in most devices e.g.s smartphones

@levelsio

261,455 views • 2 years ago

We used to buy more VRAM. Now we’re finding ways around the VRAM limit. This little DDR5 SO-DIMM setup is paired with an old NVIDIA Tesla P40 in the video. The P40 is an old datacenter GPU with 24GB of VRAM. By today’s standards, it’s not fast. But combine that with a machine carrying 128GB of system memory, and suddenly you have a very different local AI box. The GPU handles what it’s good at. The system memory gives the model somewhere to go when VRAM runs out. You obviously pay for that with bandwidth and inference speed. But this is the part I find interesting: Local AI is making “how much memory do I have?” almost as important as “how powerful is my GPU?” We’ve spent years chasing faster GPUs. Maybe the next wave of local AI hardware will be about finding smarter ways to use all the memory we already have. More VRAM is still better. But it’s no longer the only way to run a bigger model.

We used to buy more VRAM. Now we’re finding ways around the VRAM limit. This little DDR5 SO-DIMM setup is paired with an old NVIDIA Tesla P40 in the video. The P40 is an old datacenter GPU with 24GB of VRAM. By today’s standards, it’s not fast. But combine that with a machine carrying 128GB of system memory, and suddenly you have a very different local AI box. The GPU handles what it’s good at. The system memory gives the model somewhere to go when VRAM runs out. You obviously pay for that with bandwidth and inference speed. But this is the part I find interesting: Local AI is making “how much memory do I have?” almost as important as “how powerful is my GPU?” We’ve spent years chasing faster GPUs. Maybe the next wave of local AI hardware will be about finding smarter ways to use all the memory we already have. More VRAM is still better. But it’s no longer the only way to run a bigger model.

KyzoroX

35,554 views • 3 days ago

Astra built this three.js game with gorgeous graphics, then recreated every scene in Blender and made the trailer too. It feels like a massive upgrade from Sol for 3D graphics and front-end design. The UI feels way more polished and less like AI slop. The rocks, trees, monuments, and characters are all procedural code, keeping the game under 2 MB before the image and music assets. The level of autonomy is unbelievable. It needed very little steering, though I still had to keep pushing it to work harder on the details. The UI icons, buttons, and ornaments were made as transparent PNGs with GPT Image 2.

Astra built this three.js game with gorgeous graphics, then recreated every scene in Blender and made the trailer too. It feels like a massive upgrade from Sol for 3D graphics and front-end design. The UI feels way more polished and less like AI slop. The rocks, trees, monuments, and characters are all procedural code, keeping the game under 2 MB before the image and music assets. The level of autonomy is unbelievable. It needed very little steering, though I still had to keep pushing it to work harder on the details. The UI icons, buttons, and ornaments were made as transparent PNGs with GPT Image 2.

Meng To

134,358 views • 7 days ago

Not bad! 2K triangles and a bit of manual cleanup. This was a quick test creating a lowpoly 3D model. Tried Tripo for retopology and it did a good job. It does render some inverted normals though. For this lowpoly model it’s fine but inverted normals are a pain when working on highpoly models. I’m going to test some other options like the remesh tool in Blender sculpting. It’s crazy how this Trellis Texturing node works btw, It understands the geometry very well. Watch the end where I made the worst version of an Axe in Blender and it was able to wrap the texture on that model! I would probably spend a bit more time splitting the image up and generating the main parts separately. Also making any signs readable or just remove them.

Not bad! 2K triangles and a bit of manual cleanup. This was a quick test creating a lowpoly 3D model. Tried Tripo for retopology and it did a good job. It does render some inverted normals though. For this lowpoly model it’s fine but inverted normals are a pain when working on highpoly models. I’m going to test some other options like the remesh tool in Blender sculpting. It’s crazy how this Trellis Texturing node works btw, It understands the geometry very well. Watch the end where I made the worst version of an Axe in Blender and it was able to wrap the texture on that model! I would probably spend a bit more time splitting the image up and generating the main parts separately. Also making any signs readable or just remove them.

Jerome | InsaneUnreal

15,052 views • 7 months ago

🔥 This WEB3 GAME coming to APECHAIN is FIRE! 🔥 We teamed up with some content creators and gamers from Wayfinders to test CODE: Havoc , and it’s been a blast! 🔴 Here’s what stood out: - 3rd person shooter with fast-paced, fun gameplay - Tons of weapons - Runs super smooth—even on low-end PCs! (I had some latency since there are no LATAM servers yet ... TEST ) - Bet your ApeCoin in matches! 💰 This is definitely one to keep an eye on! 👀

🔥 This WEB3 GAME coming to APECHAIN is FIRE! 🔥 We teamed up with some content creators and gamers from Wayfinders to test CODE: Havoc , and it’s been a blast! 🔴 Here’s what stood out: - 3rd person shooter with fast-paced, fun gameplay - Tons of weapons - Runs super smooth—even on low-end PCs! (I had some latency since there are no LATAM servers yet ... TEST ) - Bet your ApeCoin in matches! 💰 This is definitely one to keep an eye on! 👀

EL REA 🇦🇷

10,813 views • 1 year ago

Dylan Patel of SemiAnalysis says a worse GPU with better storage and memory now beats the best chip without them, so buying the newest GPU alone no longer wins inference. So, an AMD GPU with more memory can outperform Nvidia in some cases. "So what we have is we have over $80 million of compute, GPUs from Nvidia, AMD, TPUs from Google, Trainium from Amazon, and we run this benchmark constantly on the newest inference engine, newest drivers, newest PyTorch version, whatever it is." "Every day it runs on an automated CI, and we run it on all the latest Chinese models, from GLM, Zhipu, Moonshot, Kimi, Alibaba, all these models we run." "Initially, when we were benchmarking the difference between these chips and different engines, different schemes for parallelism, we were just running it fixed context length." "But now with Agent X, we've analyzed over $5 million worth of Claude Code traces. This is real production traffic that people have donated to us as well as internally generated. Now we know what the actual agent workload looks like." "And then as we implement that and run those benchmarks, it turns out yes, the chip you're using is very important, but now even more important is how are you handling this memory offload?" "And so while an Nvidia GPU is faster than an AMD GPU in most cases, because AMD GPUs have more memory, they actually end up outperforming in some cases." "Or you can have a worse GPU, but a much better storage solution, and now you can outperform what the best GPU can do without those solutions. So just buying the newest and latest GPU alone doesn't get you the best inference economics." "Actually, you need to layer in all these other innovations including storage and memory." [ Who's the top player on your chart? ] "That really is a difficult multivariable problem. And generally that means you need to have, yes, you need to have the best GPU, a GB300, but you also need to have the best storage solutions. And so I won't spoil who's the best right here, but I will say that storage solutions matter a lot and memory solutions matter a lot, as does your front-end networking. That matters a lot."

Dylan Patel of SemiAnalysis says a worse GPU with better storage and memory now beats the best chip without them, so buying the newest GPU alone no longer wins inference. So, an AMD GPU with more memory can outperform Nvidia in some cases. "So what we have is we have over $80 million of compute, GPUs from Nvidia, AMD, TPUs from Google, Trainium from Amazon, and we run this benchmark constantly on the newest inference engine, newest drivers, newest PyTorch version, whatever it is." "Every day it runs on an automated CI, and we run it on all the latest Chinese models, from GLM, Zhipu, Moonshot, Kimi, Alibaba, all these models we run." "Initially, when we were benchmarking the difference between these chips and different engines, different schemes for parallelism, we were just running it fixed context length." "But now with Agent X, we've analyzed over $5 million worth of Claude Code traces. This is real production traffic that people have donated to us as well as internally generated. Now we know what the actual agent workload looks like." "And then as we implement that and run those benchmarks, it turns out yes, the chip you're using is very important, but now even more important is how are you handling this memory offload?" "And so while an Nvidia GPU is faster than an AMD GPU in most cases, because AMD GPUs have more memory, they actually end up outperforming in some cases." "Or you can have a worse GPU, but a much better storage solution, and now you can outperform what the best GPU can do without those solutions. So just buying the newest and latest GPU alone doesn't get you the best inference economics." "Actually, you need to layer in all these other innovations including storage and memory." [ Who's the top player on your chart? ] "That really is a difficult multivariable problem. And generally that means you need to have, yes, you need to have the best GPU, a GB300, but you also need to have the best storage solutions. And so I won't spoil who's the best right here, but I will say that storage solutions matter a lot and memory solutions matter a lot, as does your front-end networking. That matters a lot."

Fireside Alpha

178,511 views • 1 month ago

The reason why Reflections in GTA 6 look noisy and blurry isn't because Rockstar Games doesn't know how to do them or that a $2 billion budget magically fixes this. It all comes down to the GPU in current gen consoles, these consoles don't have the power they need to render high resolution Ray Traced reflections in real-time and corners had to be cut. Here I'm running GTA 5 enhanced for PC at ultra settings to show you how good and clean the reflections look if you have high-end and recent hardware that have AI upscaling to make this possible. That being said, The PS5 Pro console is much more powerful compared to the base PS5 (the footage which GTA 6 was captured) offering higher resolution, 45% faster rendering, 2X-3X better Ray Tracing performance and PSSR upscaling. With the PS5 Pro, it will have significantly better looking and sharper reflections somewhat close to what GTA 5 Enhanced reflections look like on a gaming PC at high settings. However, the ultimate experience will still be on PC when GTA 6 launches, allowing for the highest quality Ray Tracing or even Path Tracing if Rockstar Games decides to implement it.

The reason why Reflections in GTA 6 look noisy and blurry isn't because Rockstar Games doesn't know how to do them or that a $2 billion budget magically fixes this. It all comes down to the GPU in current gen consoles, these consoles don't have the power they need to render high resolution Ray Traced reflections in real-time and corners had to be cut. Here I'm running GTA 5 enhanced for PC at ultra settings to show you how good and clean the reflections look if you have high-end and recent hardware that have AI upscaling to make this possible. That being said, The PS5 Pro console is much more powerful compared to the base PS5 (the footage which GTA 6 was captured) offering higher resolution, 45% faster rendering, 2X-3X better Ray Tracing performance and PSSR upscaling. With the PS5 Pro, it will have significantly better looking and sharper reflections somewhat close to what GTA 5 Enhanced reflections look like on a gaming PC at high settings. However, the ultimate experience will still be on PC when GTA 6 launches, allowing for the highest quality Ray Tracing or even Path Tracing if Rockstar Games decides to implement it.

NikTek

142,884 views • 12 days ago

UC Berkeley just open-sourced FreeToken. (2–4x faster local LLM inference than Ollama) the results are wild: - Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s - DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s - GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s a 35B model at 16-bit precision needs about 70GB just for its weights. even at 4 bits it is close to 18GB, and FreeToken serves it on an 8GB GPU. let me explain how: all three models mentioned above are Mixture-of-Experts, and that is what FreeToken takes advantage of. each layer holds hundreds of separate experts plus a small router that picks a few of them per token. Qwen3.6-35B activates roughly 3B of its 35B parameters per token. DeepSeek-V4-Flash picks 6 of 256 experts per layer, so 13B of its 284B run at a time. so compute was never the bottleneck. the weights a single step touches fit comfortably on a consumer GPU. every expert the router might pick still has to exist somewhere. they sit in system RAM, and the GPU keeps a cache of the ones the model has been using recently. so everything comes down to what happens when the router picks an expert that is not on the GPU. there are two ways to serve that miss: 1. copy it over PCIe and run it on the GPU 2. run it on the CPU, where it already lives both read from the same system memory, so they compete for one pool of bandwidth instead of adding to each other. existing engines pick one option and freeze it when the model loads. but routing changes on every token, so a fixed choice misses most of what the model asks for. FreeToken measures both bandwidths on your machine and splits each step's misses between the two paths in proportion. the GPU and CPU results then merge exactly, with no approximation. two machines with the same GPU can end up wanting opposite strategies, which I did not expect. a 5090 in a gaming desktop should push nearly everything over PCIe, while an 8GB laptop is better off computing most misses on the CPU. none of that is readable off a spec sheet, so the engine profiles it once per machine. the second half of the design is about agents. coding agents constantly rewrite their own history, and every edit normally forces thousands of tokens back through prefill. FreeToken saves its checkpoints at the exact boundaries agent frameworks cut on, so it only reprocesses the new part. its slowest first token stays under 44 seconds, while llama.cpp peaks at 232 and KTransformers at 946. it serves the OpenAI and Anthropic APIs under Apache 2.0, so Claude Code and Codex can point at it directly. releasing weights publicly decides who can download a model, not who can afford to run one. frontier open models keep shipping, and running them still assumes a rented cluster. meanwhile there are over a hundred million consumer machines with discrete GPUs sitting mostly idle. closing that gap was never a hardware problem, and work like this is what turns open weights into something you can actually use. paper: repo: almost every idea in this post, from why memory bandwidth decides the outcome to why moving weights costs more than computing on them, comes straight out of how a GPU is built. I wrote a detailed primer on that. the article is quoted below.

UC Berkeley just open-sourced FreeToken. (2–4x faster local LLM inference than Ollama) the results are wild: - Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s - DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s - GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s a 35B model at 16-bit precision needs about 70GB just for its weights. even at 4 bits it is close to 18GB, and FreeToken serves it on an 8GB GPU. let me explain how: all three models mentioned above are Mixture-of-Experts, and that is what FreeToken takes advantage of. each layer holds hundreds of separate experts plus a small router that picks a few of them per token. Qwen3.6-35B activates roughly 3B of its 35B parameters per token. DeepSeek-V4-Flash picks 6 of 256 experts per layer, so 13B of its 284B run at a time. so compute was never the bottleneck. the weights a single step touches fit comfortably on a consumer GPU. every expert the router might pick still has to exist somewhere. they sit in system RAM, and the GPU keeps a cache of the ones the model has been using recently. so everything comes down to what happens when the router picks an expert that is not on the GPU. there are two ways to serve that miss: 1. copy it over PCIe and run it on the GPU 2. run it on the CPU, where it already lives both read from the same system memory, so they compete for one pool of bandwidth instead of adding to each other. existing engines pick one option and freeze it when the model loads. but routing changes on every token, so a fixed choice misses most of what the model asks for. FreeToken measures both bandwidths on your machine and splits each step's misses between the two paths in proportion. the GPU and CPU results then merge exactly, with no approximation. two machines with the same GPU can end up wanting opposite strategies, which I did not expect. a 5090 in a gaming desktop should push nearly everything over PCIe, while an 8GB laptop is better off computing most misses on the CPU. none of that is readable off a spec sheet, so the engine profiles it once per machine. the second half of the design is about agents. coding agents constantly rewrite their own history, and every edit normally forces thousands of tokens back through prefill. FreeToken saves its checkpoints at the exact boundaries agent frameworks cut on, so it only reprocesses the new part. its slowest first token stays under 44 seconds, while llama.cpp peaks at 232 and KTransformers at 946. it serves the OpenAI and Anthropic APIs under Apache 2.0, so Claude Code and Codex can point at it directly. releasing weights publicly decides who can download a model, not who can afford to run one. frontier open models keep shipping, and running them still assumes a rented cluster. meanwhile there are over a hundred million consumer machines with discrete GPUs sitting mostly idle. closing that gap was never a hardware problem, and work like this is what turns open weights into something you can actually use. paper: repo: almost every idea in this post, from why memory bandwidth decides the outcome to why moving weights costs more than computing on them, comes straight out of how a GPU is built. I wrote a detailed primer on that. the article is quoted below.

Akshay 🚀

340,691 views • 21 days ago