0xSero

@0xSero • 75,097 subscribers

do you think like I do?

Shorts

17,260 views

122,193 views

69,259 views

135,941 views

21,334 views

37,219 views

22,222 views

22,535 views

22,762 views

13,869 views

24,131 views

11,809 views

Videos

sweetdream.ai

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Private Show

Join now for exclusive access

Free preview available • Premium content

191,423 views • 18 days ago

29,762 views • 3 days ago

92,222 views • 10 days ago

186,914 views • 29 days ago

99,299 views • 16 days ago

18,526 views • 3 days ago

17,212 views • 4 days ago

201,663 views • 1 month ago

33,845 views • 11 days ago

175,626 views • 2 months ago

104,452 views • 1 month ago

56,977 views • 24 days ago

38,607 views • 17 days ago

65,627 views • 1 month ago

92,814 views • 1 month ago

145,527 views • 2 months ago

24,760 views • 13 days ago

43,765 views • 24 days ago

158,188 views • 3 months ago

18,820 views • 12 days ago

Live Cam

0xSero

Shorts

This is running on a 16gb mac mini and 3090 Incredible model beats most of the mid stuff, qwen3.8-next-flash in exl3-3bits over half the model is being streamed from NVME it's 31 tok/s that is faster than Codex as of late.. for 2600$ I feel like my mission needs updating

Tibo, why? Why must the performance suck so bad.

100 tok/s on GLM-5.3-Flash, GLM-5.3 runs on the 6000s @ 80-180 tok/s I have not tried the Studio, I’m sure it’s an amazing machine especially if you are in the Mac world already. From a pure logistics POV the sparks use little power, make less noise, serve good model &amp; tokens

Codex nooooooo I can't close this popup

Local AI: - can see reasoning - faster - will do whatever I want

Deepseek-v4-Flash running locally. Makes me dizzy, this is one of 6 sessions running rn

Kitty litter, GLM-5.2 running locally. Have you seen 180 tok/s? Do you see how much room we have to push things forward? Most owners don’t know how to squeeze every token out of their GPUs

GLM-5.2-REAP-NVFP4 This experiment I calibrated the model on 30,000+ samples of my agent sessions, tweets, writing, and codebases. Pretty charming, every dataset produced something different.

The fact that a 284B parameter model is able to in a single 400K token session - ssh into my Homelab to find docs - ssh into lambda / prime intellect - rewrite reap to support the DS4 attention - download models &amp; datasets - run tests to check it works

If you've done work in OSS AI, or Local AI and want to have you and your company on State of Local AI - 2026 Get in touch, you'll be on as a coauthor with the companies and people who make this all possible. We moved release date to August 7th, we have 6 contributors so far

Asking Codex (GPU) to find and kill Codex (Cerebras)

What 4 concurrency on GLM sounds like. So gosh damn loud, open air frames are easier to build but so gosh darn loud

Videos

Watch Anya Live

Running a local agent 24/7 until it successfully gets paid for a security bug bounty Qwen3.8-27B - intel arc b70 - 262k context - vision - 60 tok/s decode - 1500-6000 tok/s prefill

Guess the model, it took about 2 hours. "Make me a full 3d render in blender and godot of San Francisco."

If you’re someone who has 1 or more expensive GPUs and you want to run bigger models without dropping 5 digits to get more cards 1. Buy M2 &gt; PCIe gen 5 raid0 adapter 2. Buy 4x 1TB Samsung gen 5 NVMe You can get 60 GB/s on these fast enough to move data in and out of GPUs

I interviewed DHH, the creator of Omarchy - The Agent Operating system + Ruby on Rails, one of the most important frameworks of our time. He's a father, builder, race-car driver, and such an inspiring person. One of the most positive people I've had the honor of meeting.

You thought I was joking 🤓 Here's Deepseek-v4.1-Flash running on CPU + DDR4 with just a single RTX 3090 Wow Mia great quant (:

You think you could go back to 20 tok/s after this?

Qwen3.8-next-flash has been going for 5 hours!!! on 3x DGX Spark. Prompt: make a photorealistic 3d version on San Francisco. It's got drone views, skateboarding, driving and junping around the map. GLM-5.3 and Flash were both done in an hour and put out garbage wtf

Opus at home, at 200+ tok/s I love ZAI

I read Pi because the code is good, it's well thought through, it's simple, it works. Armin, and Mario share their experience building Pi, and how they're thinking through their roadmap. Such an honour to sit down with amazing engineers (:

I just saw AGI, what the heck. Deepseek-v4-flash-0731 invented vision for itself, this is so incredible. Result in the next post:

3 months ago I said local AI will be competitive.

What local ai looks like at 2000$ Left: qwen3.8-27b in 4 bits on 1x 3090 w vision and full context right: same model running on intel arc b70

Explaining what I think I understood about NVMe-PCIe adapters, I’m hoping i can make expert offloading for qwen3.6-35b or qwen3.8-flash-next working Might be able to get away with 8-12GB of VRAM although I’m just saying things at this point This bring down the cost to run tons

I told you Local AI will be competitive. 12 / 15 sessions running on my machine rn are local, DS4.1-Flash and GLM-5.3-Flash are better than anything else out there. So fast, don't complain, no rate limits. Why we gotta live like plebs?

Learn about Harnesses! I went over all these - omp - pi - claude code - zcode - codex - warp This is what I am using everyday, I explain the difference and benefits of each.

Yours to keep.

Harness tierlist TLDR: Deepseek and Pi rock, antigravity is really bad

This is the best thing I've ever designed. I am so happy with how it turned out, I will be using my b70s Omarchy turns your pc into AI Free AI that works, battle tested - codex - claude code - pi / omp - crush - opencode If you have 12GB of Nvidia VRAM build a box and join

GLM-5.2 Nvidia NVFP4 No pruning, no quantising, exact same model card. 110 tok/s single stream. My friend is a genius, holy fudge. Here's the repo, he got full DS4-Flash on a 5090 + DDR5 at 38.5 tok/s This is a breakthrough.

Deepseek-v4.1-flash can build 3d worlds like no other local model, it's incredible at this kind of task. 1 shot! 40 score on AA for this model is just wrong, it's so good. It's just weak in UI, and has about 200M token window before it goes off track

100 tok/s on GLM-5.3-Flash, GLM-5.3 runs on the 6000s @ 80-180 tok/s I have not tried the Studio, I’m sure it’s an amazing machine especially if you are in the Mac world already. From a pure logistics POV the sparks use little power, make less noise, serve good model & tokens

The fact that a 284B parameter model is able to in a single 400K token session - ssh into my Homelab to find docs - ssh into lambda / prime intellect - rewrite reap to support the DS4 attention - download models & datasets - run tests to check it works

If you’re someone who has 1 or more expensive GPUs and you want to run bigger models without dropping 5 digits to get more cards 1. Buy M2 > PCIe gen 5 raid0 adapter 2. Buy 4x 1TB Samsung gen 5 NVMe You can get 60 GB/s on these fast enough to move data in and out of GPUs