OpenAI Dots & GrokBot alternative running fully locally!! It uses PolyStrata (Strata engine) to serve and swap the models it needs and, of course being fast 😎
Check out more here: github.com/simplexai-labs/li…
Vik | Practical AI & Tech retweeted
what the f*ck, you can now run uncensored AI on a 6 year old laptop with no GPU 😳
Someone made an uncensored build of Google's Gemma 4 E4B, and it's only 4.8GB
- runs on 8GB of RAM
- works on an old GTX 1060 or 1650
- refuses way less than normal Gemma
- Apache-2.0, free forever
- runs offline, so nothing leaves your laptop
set it up in 2 mins:
→ install Ollama
→ open your terminal
→ paste: ollama run hf.co/igorls/gemma-4-E4B-it-…
→ start chatting
most of you don't have a 16GB card, this is the one to try first.
try it here 👇
huggingface.co/igorls
jan is the free offline ChatGPT big tech hates you knowing about 😳
45.7K stars on GitHub, fully open source
- runs AI models on your own pc, no internet needed
- download Llama, Gemma, Qwen and more right inside the app
- works on windows, mac and linux
github.com/janhq/jan
save this and stop paying $20 a month
Strata is also now natively integrated in @atomic_chat_hq ! 🥳
See more at: github.com/AtomicBot-ai/Atom…
Introducing Atomic Local AI Server ⚡
Choose inference engine. Run local models. Connect your agents.
• Multiple engines: llama.cpp, vLLM, SGLang, TensorRT, MLX, Prism and more.
• Local AI server with an OpenAI-compatible API.
• One-command engine installation matched to your GPU, CPU or NPU.
• Pull from 3M Hugging Face models in GGUF, MLX or safetensors formats.
• Run on Windows, Linux, Apple Silicon and NVIDIA hardware, including RTX, Blackwell and DGX Spark. AMD Ryzen Halo support coming soon.
• Connect Hermes, OpenClaw, Claude, Codex, Pi agent and others.
• Generate images and videos with stable-diffusion.
• Manage everything through CLI or web UI.
• Remote access via Cloudflare tunnels
Free. Open source. Local. 👇
github.com/AtomicBot-ai/atom…
Vik | Practical AI & Tech retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
Vik | Practical AI & Tech retweeted
It started as a weekend project to learn local inference. Now, it's a growing community.
Thank you for 1,000 GitHub stars and 100+ forks!
A special thanks to all the contributors and everyone who spent time testing and sharing ideas.
This is just the beginning!
Bit by bit
🚀 Strata v0.2.14 is out! Released today, November 15, by @coldniko (Niko Veit).
This is a major architectural milestone: Strata now runs prompt-expert computation on its own native GPU kernels, replacing the previous llama.cpp path on NVIDIA and AMD GPUs (Windows/Linux).
And the benchmarks are impressive:
🔥 RTX 5090: 7,870 vs 6,851 tok/s — +14.9% prompt processing
🔥 RTX 4090: 20K-token prompt in 5.65s vs 6.11s — ~7.5% faster
🔥 RTX 5070 Ti: 1,408 vs 953 tok/s — +47.7% on a 20K-token prompt
🔥 Dual RTX A6000: 75.88 vs 64.16 tok/s — +18.3% output with improved multi-GPU memory allocation
🔥 Mac M5 Max: 94.7 vs 92.8 tok/s output with IQ3_S.
No new settings. Existing models, chats and configurations remain intact.
What makes this release particularly interesting isn’t just the speed improvement. Strata is progressively replacing inherited inference components with its own implementations.
An independent inference engine is taking shape, release by release.
Huge respect to Niko for the pace of development. We’re looking forward to testing this version on our own hardware!
👨💻 Niko:
🔗 GitHub release: github.com/Niko1221/Strata/r… nitter.cf/coldniko
I’ve worked that heatshield my entire adult life. Peak career moment today
To go from little 7 tile patches to a full reusable ship. Absolute marathon
I’m so grateful I got to do something useful, that I was given the chance
I don’t really have many good memories from this experience.
It was so difficult. It almost looks easy now, but none of it came easily
Please do something amazing with this rocket. This tech is such a gift, a love letter to the species
Vik | Practical AI & Tech retweeted
Here's a budget localmaxxing find. I've been keeping this one to myself because I don't want it to sell out.
Could be a fun holiday gift for someone who wants to start experimenting with local AI.
This is the Origimagic C7 mini PC - ~30% off.
I paid $289 for 16GB DDR4, a 512GB SSD, and two Thunderbolt 4 ports.
I added another 16GB for $89, bringing mine to 32GB RAM for $378 total, before the eGPU.
It’s the only cheap Thunderbolt 4 PC I’ve found that checks my boxes, and it works with at least one Gigabyte eGPU.
Next experiment - seeing whether I can run two eGPUs on this thing.
Currently $299 for the base machine. Add another 16GB for $100
Anyone else running a mini PC + eGPU localmaxxing setup?
Affiliate links in the thread.
Vik | Practical AI & Tech retweeted
😯Why not buy a $702 AMD GPU from 2018 (~1TB/s memory bandwidth) with more VRAM than a $1,790 RTX 3090?
And it's running Qwen3.8-27B at 8-bit precision.
A local AI enthusiast has been tracking used AI GPU prices on eBay through a new tool called RackBites.
Here are the most recent numbers.
GPU → VRAM → Low-end asking price
🟢 Tesla P100 → 16GB → $174
🟢 Tesla P40 → 24GB → $280
🟢 Tesla V100 → 16GB → $220
🔴 AMD MI50 → 32GB → $702
🟢 Tesla V100 → 32GB → $949
🟢 Quadro RTX 8000 → 48GB → $2,111
🟢 RTX 3090 → 24GB → $1,790
🔴 Radeon AI PRO R9700 → 32GB → $2,430
But the AMD MI50 really caught my attention.
🔴 MI50 32GB
💾 HBM2 memory
⚡ ~1TB/s memory bandwidth
💵 $21.93 per GB of VRAM
And people are still optimizing local AI software for it.
Check out a recent single-GPU benchmark using a specialized llama.cpp fork
🧠 Qwen3.8-27B
💾 Q8_0 weights + Q8_0 KV cache
⚡ 2K context → 22.8 tps
⚡ 64K context → 20.4 tps
🔥 64K reasoning + MTP → 28.8 tps
💻 Short coding + MTP → 51.8 tps
On a GPU released in 2018!
Now here's another interesting detail.
The RackBites developer bought an 8× V100 32GB server in early September for $5,400.
256GB of GPU memory.
The same server is now reportedly listed for $13,000.
Are older datacenter GPUs becoming more valuable as Localmaxxers find new ways to run modern models on them?
These are eBay asking prices, not completed sales.
Links in replies
Vik | Practical AI & Tech retweeted
You know how I work with vulns at my job? Let me tell you.
I run dozens of subagents on DeepSeek V4.1 Flash Abliterated. Each one pushes 350-400 tok/s. All of it is self-hosted on our own bare metal.
And it is only dozens because I do not have more vuln tasks than that. Give me more work and I just start more agents. The hardware is not the limit.
Why abliterated? Because to a safety filter, security work looks exactly like an attack.
When Hugging Face got hit by an AI agent attack this year, its security team tried the top commercial models first. The models refused to look at the attack.
Why self-hosted? Our code and our findings never leave our servers, and nobody can rate limit us in the middle of a scan.
That is 1 guy with a small queue of tasks.
Now imagine what the big players are running.
State teams. Big security companies. People who never post about it.
They are not waiting for the debate about which models Hugging Face should delete.
Vik | Practical AI & Tech retweeted
Big new open-weights model! This one could be a top model. Check out Code Arena scores.
Beats DeepSeek V4.1 and Kimi K3 in coding
Baidu released a 280B MoE model that activates only 14B parameters per token.
And it's already challenging DeepSeek and Kimi in coding.
🇨🇳 This is ERNIE 5.5 Flash Preview.
Model Stats
🧠 280B total / 14B active
📚 1 MILLION token context
⚡ Baidu reports 200+ tokens/sec on its hosted platform
🔧 Coding, tool calling and agentic workflows
🔓 Apache 2.0 weights coming soon
Now look at the early Code Arena WebDev scores:
🔥 ERNIE 5.5 Flash Preview → 1668
🟢 DeepSeek V4.1 Flash (Max) → 1664
🔵 Kimi K3 (Max) → 1658
These are early results, with ERNIE's score based on Arena's automated evaluation rather than human votes.
Baidu is showing ERNIE working across entire agent workflows like these examples ...
🎮 Built a third-person Unreal Engine 5.8 game in an 11-hour coding session.
🕹️ Built nearly 100 mobile mini-games from design documents.
✈️ Two agents collaborated to build an airport and Boeing 747 in Blender.
⚙️ It created an editable W16 engine in FreeCAD.
📊 Produced research reports, spreadsheets and presentations.
How much hardware will we need?
Raw weight math for 280B parameters:
💾 BF16 → 560GB
💾 8-bit → 280GB
💾 4-bit → 140GB
💾 3-bit → 105GB
💾 2-bit → 70GB
Before quantization metadata, KV cache and runtime overhead.
⚠️ API available now. The post-trained Preview weights are NOT downloadable yet.
Baidu says Apache 2.0 weights are coming.
Link in ALT
Vik | Practical AI & Tech retweeted
I tested 8 AI models on 12 questions from my own field. Every question 3 times.
Textbook questions: 96 % correct.
My real paperwork: 4 %.
Times a model said "I don't know": 0 of 288.
Full results: os-ai.thomasbaertl.com/not-o…
Vik | Practical AI & Tech retweeted
I built an X-ray mirror using Claude Fable 5.2. 💀
Point a webcam at your face and see the skull, brain and muscles inside your head, tracked live as you move. Open your mouth and the jaw opens. Hold up a hand and you see its bones.
One web page, no install, and the video never leaves the computer. Built with Claude Fable 5.2 by describing what I wanted. You can download it and try it for free. Link in the comment section. 👇🔗
Anatomy models: BodyParts3D, © DBCLS, CC BY-SA 2.1 Japan.
Vik | Practical AI & Tech retweeted
Apple is working on a new app called Harbor
Harbor turns your Mac into a private AI inference server. Run open-weight models locally on Apple silicon—your data never leaves your hardware, and nothing reaches the cloud. Connect multiple Mac computers over Thunderbolt 5 to pool unified memory, enabling large models and higher concurrency that a single computer can handle alone.
Set up in minutes, not hours Harbor automatically detects your devices over Thunderbolt 5 to form a cluster. Plug in, start, and you’re running—no manual configuration, no command line, no infrastructure expertise required.
Download and run models directly Browse, download, and load open-weight models from within the app. Harbor handles model management so you can focus on the work, not the setup.
Cluster for scale Connect two or more Thunderbolt 5-capable Mac computers and Harbor pools their unified memory into a single inference resource. Run models that exceed any single computer’s capacity, or serve more concurrent requests across your cluster.
Harbor is currently in a private preview under the Claris brand.

































