João Correia (@Pipolinus) header
João Correia (@Pipolinus) profile picture
Joined May 2009
Photo by João Correia (@Pipolinus)Photo by João Correia (@Pipolinus)Photo by João Correia (@Pipolinus)Photo by João Correia (@Pipolinus)Photo by João Correia (@Pipolinus)Photo by João Correia (@Pipolinus)Photo by João Correia (@Pipolinus)Photo by João Correia (@Pipolinus)Photo by João Correia (@Pipolinus)Photo by João Correia (@Pipolinus)
João Correia retweeted
Replying to @superalesha
Grabbing every file in these repos is ~13.3 TB. A more realistic option: one quant per model that runs on an RTX 4090 (24GB VRAM, 64GB RAM). ~279 GB total. LLMs - Qwen3.8-27B UD-Q4_K_M (16.6 GB): hf.co/unsloth/Qwen3.8-27B-GG… - Qwen3.8-27B abliterated Q4_K_M (16.6 GB): hf.co/huihui-ai/Huihui-Qwen3… Image - Qwen-Image 2.1 bf16 (33.1 GB): hf.co/Qwen/Qwen-Image-2.1 - MOSS-Image Base (27.2 GB): hf.co/OpenMOSS-Team/MOSS-Ima… - FLUX.3 Klein int8 + int8 encoder + VAE (11.0 GB, gated): hf.co/black-forest-labs/flux… Video - Wan 2.2 T2V + I2V A14B fp8 high/low + umt5 fp8 + VAE (64.2 GB): hf.co/Comfy-Org/Wan_2.2_Comf… - HunyuanVideo 1.5 720p t2v distilled + VAE (38.4 GB): hf.co/tencent/HunyuanVideo-1… - LTX-2.5 distilled int8 + int8 encoder + VAEs + upscaler (39.7 GB, gated): hf.co/Lightricks/LTX-2.5 Voice / TTS - Fish Audio S2 Pro (11.0 GB): hf.co/fishaudio/s2-pro - Qwen3-TTS 1.7B Base (4.5 GB): hf.co/Qwen/Qwen3-TTS-12Hz-1.… Music - ACE-Step 1.5 (10.1 GB): hf.co/ACE-Step/Ace-Step1.5 - YuE2 int8 + audio encoder (5.4 GB): hf.co/Comfy-Org/YuE2 Speech to text - Whisper Large v3 Turbo (1.6 GB): hf.co/openai/whisper-large-v…
1
1
1
33
2,158
João Correia retweeted
DOWNLOAD THESE 24 OPEN MODELS BEFORE MORE REPOS GET GATED OR TAKEN DOWN. You asked for the exact links, so I put together the full local AI pack: LLMs, uncensored quants, image generation, video, voice, music and speech to text. Put them on your own disk while you still can. 2 repos in this list already block anonymous downloads behind a login gate, and Hugging Face has disabled uncensored repos with zero warning before. If a provider bans your account, raises prices or rewrites its rules next month, nobody can take away files that are already on your drive. For all 4 LLMs below I included 3 versions: original weights, standard GGUF and abliterated GGUF with refusals removed. For Qwen3.8 Flash-Next and GLM-5.3-Flash, the uncensored GGUF repos are my own quants. [LLMs: original, GGUF and uncensored] Qwen3.8-27B Original: huggingface.co/Qwen/Qwen3.8-… GGUF: huggingface.co/unsloth/Qwen3… Abliterated GGUF: huggingface.co/huihui-ai/Hui… Qwen3.8 Flash-Next Original: huggingface.co/Qwen/Qwen3.8-… GGUF: huggingface.co/unsloth/Qwen3… Abliterated GGUF (my Strata quants): huggingface.co/alesha-pro/Qw… GLM-5.3-Flash Original: huggingface.co/zai-org/GLM-5… GGUF: huggingface.co/unsloth/GLM-5… Abliterated GGUF (my Maya-M quants): huggingface.co/alesha-pro/GL… DeepSeek V4.1 Flash 1017 Original: huggingface.co/deepseek-ai/D… GGUF: huggingface.co/unsloth/DeepS… Abliterated GGUF: huggingface.co/huihui-ai/Hui… [Image generation] Qwen-Image 2.1: huggingface.co/Qwen/Qwen-Ima… MOSS-Image Base (Apache 2.0): huggingface.co/OpenMOSS-Team… FLUX.3 Klein for ComfyUI (gated, accept terms on HF first): huggingface.co/black-forest-… [Video generation] Wan 2.2 Text-to-Video A14B (Apache 2.0): huggingface.co/Wan-AI/Wan2.2… Wan 2.2 Image-to-Video A14B (Apache 2.0): huggingface.co/Wan-AI/Wan2.2… HunyuanVideo 1.5: huggingface.co/tencent/Hunyu… LTX-2.5 (gated, accept terms on HF first): huggingface.co/Lightricks/LT… [Voice and TTS] Fish Audio S2 Pro (non-commercial research license): huggingface.co/fishaudio/s2-… Qwen3-TTS 12Hz 1.7B Base (Apache 2.0): huggingface.co/Qwen/Qwen3-TT… [Music generation] ACE-Step 1.5 (MIT): huggingface.co/ACE-Step/Ace-… YuE2 ComfyUI bundle (non-commercial CC-BY-NC-4.0): huggingface.co/Comfy-Org/YuE… [Speech recognition] Whisper Large v3 Turbo (MIT): huggingface.co/openai/whispe… [How to download cleanly] Use the official Hugging Face CLI: pip install -U huggingface_hub hf download OWNER/REPO --local-dir ./MODEL Do not pull every quant inside a GGUF repo by mistake. Pick 1 quant (and all of its split shards if it is split) with a file filter: hf download OWNER/REPO --include "*Q4_K_M*" --local-dir ./MODEL For original weights and media models, save the configs, tokenizers, text encoders, VAEs, README and LICENSE files too. For the 2 gated repos, accept the terms on the model card and run hf auth login first. An API account can get cut off in 5 seconds. Weights on your own disk stay yours even if the lab changes its mind tomorrow. Save this post and go fill your drives while every link still works.
I know what you need to do right now. Go download the weights of these 4 open models and put them on your disk. Right now, even if you cannot run the big ones yet. Nobody knows where any of this goes next. Qwen3.8-27B Qwen3.8 Flash-Next GLM-5.3-Flash DeepSeek V4.1 Flash 1017 Grab the abliterated quants too while every link still works. RIGHT NOW!
131
511
41
4,415
303,533
Are you tired of Haaland perma-cap, xMins, swingy chips, rolling free transfers, robotic solver use and teams all looking the same?! Remember when fantasy football was a bit more… fun?! Me too. And I decided to do something about it 👇
19
21
17
147
32,810
João Correia retweeted
https://teapawt.net/t.co/J40BHNPqa9 Gianluca Prestianni 2.0 | SL Benfica | 2026/27 - El Huracán - Golos, Assistências & Skills
30
4
148
14,838
João Correia retweeted
𝑆𝑡𝑎𝑡𝑠 𝑜𝑓 𝑆ℎ𝑖𝑡ℎ𝑜𝑢𝑠𝑒𝑟𝑦 𝑝𝑟𝑒𝑠𝑒𝑛𝑡𝑠 𝐓𝐇𝐄 𝐖𝐎𝐑𝐋𝐃 𝐂𝐔𝐏 𝐎𝐅 𝐌𝐎𝐕𝐈𝐄 𝐅𝐑𝐀𝐍𝐂𝐇𝐈𝐒𝐄𝐒 64 Teams. 16 Groups and some absolutely BRUTAL ones right out of the gate What is the BEST Movie Franchise of all time? LET'S DO THIS (RT far and wide)
10
21
5
55
22,588
João Correia retweeted
The old Championship Manager has been released as a browser game, it has been updated with 26/27 squads. champman2627.co.uk/play
169
788
155
8,512
1,625,223
João Correia retweeted
Every FPL manager has agonised over a benching decision, and then end up starting the blank and benching the haul. So does rotation actually work? We checked @SolioAnalytics's deadline-locked projections for every 2025-26 gameweek against what actually happened 🧵
4
21
1
108
18,443
João Correia retweeted
I built an operating system where the AI agent isn’t another app. It’s part of the OS. Today, Herald OS is open source and you can install it yourself 🚀 It’s built around Hermes Agent by Nous Research. Talk or type to Hermes and it can work across your entire computer: • Open and use apps • Find, organize and work with your files • Remember what matters • Run routines and automations • Build software while you watch • Understand what’s happening on your screen • Work across the machine instead of being trapped inside a chat window And when something is about to make a change, you approve it first. 🎙️ Talk to your computer Say “Hey Hermes” and just talk. Native OS commands don’t need to burn model tokens. When you actually ask the agent to think, build or perform a task, that’s when Hermes gets involved. ⚡ Everyday superpowers • Select any part of your screen and ask Hermes about it • Pick a colour from anywhere • Read a QR code from your screen • Copy text from anywhere on screen • Emoji in any app you type in • When an app crashes, Hermes explains what happened in plain English • One-click installs for Claude Code, Codex, local models and more 🐧 And yes, Herald OS Linux is an actual operating system • Fedora underneath, Herald on top • OS and Herald update together as one system • Roll back an update if something goes wrong • Wi-Fi, Bluetooth, sound, display and battery controls • Screen recording • Night light • Fingerprint and security-key sign-in • Firewall closed by default 🎨 Make the entire OS yours Themes don’t just change a wallpaper. They can restyle the whole system, including the terminal. Drop in any photo and build a theme from it. Or describe what you want and let Hermes design one. You can also: • Ask Hermes to create a widget, then choose exactly what that widget is allowed to access • Drag the menu bar into the layout you want • Create your own menu entries • Put your own logo in About 💻 Run it almost anywhere • Native Mac app for Apple Silicon • PC installer from a USB stick • Arch Linux package • Omarchy integration where Herald follows your theme and adds itself to the menu and keybindings • Or boot the full Herald OS inside a VM on your Mac with one command It’s free, open source and MIT licensed. It’s also still alpha, so expect rough edges. This is an independent project and is not affiliated with Nous Research. I’ve been building toward this idea for a while: Instead of opening an AI app on your computer, what if your computer itself understood what you wanted? Now you can try it ⭐ github.com/iamlukethedev/Her…
Didn't want to say anything but I kinda started something a few days ago
138
113
44
1,090
116,747
João Correia retweeted
Introducing Project Maya 🟣 GLM-5.3-Flash - a 321B parameter model - running on your own GPU. ⚡ Up to 40 tokens/s answering, up to 440 tokens/s reading your prompt (2x Tesla V100 + 30 GB RAM) 🧠 The whole 321B model in a 90 GB quant made from Z.ai's own FP8 weights 📏 Steady speed as the conversation grows: Maya keeps its pace with context depth - a 60K-token conversation answers about as fast as a short one. How it works: Maya's engine keeps the most-used experts on your GPU, the next ones in RAM and the rest on your SSD, and moves them as the conversation needs them. It adapts to whatever hardware it finds - one GPU or two, more or less RAM. It runs with 32 GB of RAM; more memory is faster. What you get: • 🖼️ Pictures - the vision encoder only loads while it reads a new image, so the model keeps the whole GPU the rest of the time • 🔌 OpenAI- and Anthropic-compatible API - plug in your apps and coding agents • 🚀 MTP speculative decoding with the model's own draft layer • 🧩 Our own quants (Maya-S), measured against the full FP8 model • 💬 A chat dashboard with live monitoring of the model and your hardware • 🛠️ One-command install: git clone, ./setup.sh, answer a few questions - it compiles the engine for your GPU, downloads the model (checksum-verified) and starts Every machine is different - try it on yours and tell us what you get. Built on Strata (MIT) and ggml/llama.cpp. 🙏
39
60
16
466
42,026
João Correia retweeted
Drake Maye to Efton Chism: Truly Art 🙏
50
480
38
4,430
202,922
João Correia retweeted
À atenção de Miguel Pinto Luz, Carlos Moedas e Pedro Duarte.
16
17
5
65
6,471
João Correia retweeted
oh btw, omp can configure itself starting next update!
22
9
287
9,277
João Correia retweeted
Zuckerberg admitted it on stage last night. Meta's personal agent gets paid every time it spends your money. Muse launched three weeks ago. It's an agent that lives in an app, holds your logins, and shops, books and pays on your behalf. It reached 560,000 daily users in 11 days and went to number one on both app stores. At Meta Connect, Zuckerberg said: "We believe that Muse will make you money. And we are standing behind this by making Muse free for a huge number of tokens with the expectation that over time we will profit by taking a small fee from transactions." So you use the agent for free, and Meta earns a fee on each purchase the agent makes. Three days ago Amazon blocked Muse from its store. Last night Zuckerberg announced who let it in: Walmart, Best Buy, Gap, Sephora, Wayfair, American Eagle, Expedia and Instacart, with Shopify, Stripe and PayPal handling checkout. His post about the Shopify deal put it in six words. "Shoppers find more. Shops sell more." Price comparison sites and influencers run on the same model. They take a referral fee when you buy. The difference with Muse is access. It holds your passwords and your card, and it has permission to act while you're away. Muse also got a face, a body and a voice last night, so you can video call it. It got its own email address. It got a mode on your Mac where it keeps working after you walk away, and a wake word on Meta's glasses so it can listen all day. Last month I wrote that handing an agent your passwords is insane. Last week I wrote that an agent you rent goes where its owner allows. This is the third piece. An agent that earns money when you spend is working for the shops and you are the customer it delivers. There's a version of this with no business model. An open model on your own machine, with your own logins, earns nothing from your purchases, so it has no reason to push one. It's slower but at least it answers to you. Meta built a salesman and gave it your credit card!
Here's everything I announced at Meta Connect today 👇
14
9
1
52
11,580
João Correia retweeted
Every company is building AI agents right now. Almost none of them are building the harness around them. 10 things a harness does that a model never will: (Save this before your next vendor demo.)
44
40
4
356
86,391
João Correia retweeted
1080ti, the emperor of GPUs, time to rise again 237 tok/s short — 160 tok/s long humaneval Full 128k context Everything you want to know in recipe below Had to sacrifice a GPU in the process and many sleepless nights but I’m a man of my word.
20
10
119
9,171
João Correia retweeted
Set Piece monsters! (SetPlay xG vs number of set piece deliveries received, excluding penalties) Muharemovic, Vuskovic, and Thomas proving that proactive defenders are proactive everywhere, not just when defending. Isak is surprisingly leading the chart. Follow for more! #FPL
15
6
1
361
35,241
João Correia retweeted
Botman, Thiaw and Hall have made 22 dangerous ball losses, by far the most in the league. The players to target who have shot immediately after a dangerous ball win or positional recovery from those zones over the next 6 GWs: Tchaouna Manzambi Khalaili Bobb Odegaard #FPL
9
17
1
450
36,441
João Correia retweeted
This is crazy now. github.com/ikawrakow/ik_llam… for the win!! Figures on a single 3090. Still room for improvement
22
6
1
152
22,289
João Correia retweeted
Ahead of the Champions League draw later, I built something to help football fans find cheap flights to European destinations. It's a website which lets you search by football team, and then allows you to pick a radius around start and end destinations. passmode.shop/fzzy/
96
554
103
6,241
1,854,936
João Correia retweeted
Replying to @DJLougen
I'm not finished yet. After I finish my next wave of magic I want to get it usable for 64gb RAM bros.
2
1
8
75