Founder of Colibrì a tiny engine for immense model for affordable hardware. 🐦 https://nitter.cf/t.co/QcE2sGOMBC
Joined January 2016
- Tweets1.4K
- Following765
- Followers1.6K
- Likes6.7K
Pinned Tweet
colibri v2.1.0: huge open models on your own machine. CPU, Vulkan, CUDA.
Every engine runs GGUFs natively, @UnslothAI's included. New: @MistralAI Large 4 (1T), @AIatMeta Muse Spark Open, @Recursive_SI Flint, @liquidai LFM3-VL. 256k contexts, new chat UI.
github.com/JustVugg/colibri/…
This is our new release with Qwen3.8-Flash-Next running on a local GPU in the new chat UI. You can see how fast it replies and watch the map light up as different experts activate for each token.
github.com/JustVugg/colibri/…
Thanks @fahdmirza for the tutorial!Cool video!
🐦 Colibri 2.1 is out, and I tested it hands-on
⚡ Big open models on the machine you already have, now with GGUF support and a new chat UI
🔹 One-step install, Qwen3.6-35B-A3B at 88.6 tok/s on my RTX A6000
🔹 Run a GGUF file as published, no conversion needed (Flint)
🔹 New chat UI with a live monitor and profiling page
🔹 System One decisions and the expert map
🔹 Tried a 91 GB DeepSeek V4 Flash model too, and shared what actually happened
🎬 Watch the full video below 👇
youtu.be/EdpT617DTRI?si=U2sQ…
JustVugg retweeted
🐦 Colibri 2.1 is out, and I tested it hands-on
⚡ Big open models on the machine you already have, now with GGUF support and a new chat UI
🔹 One-step install, Qwen3.6-35B-A3B at 88.6 tok/s on my RTX A6000
🔹 Run a GGUF file as published, no conversion needed (Flint)
🔹 New chat UI with a live monitor and profiling page
🔹 System One decisions and the expert map
🔹 Tried a 91 GB DeepSeek V4 Flash model too, and shared what actually happened
🎬 Watch the full video below 👇
youtu.be/EdpT617DTRI?si=U2sQ…
If you like my work on Colibri and want to support the project, buy me a coffee! Your support helps me keep optimizing the engine and dedicating more time to development.
buymeacoffee.com/justvugg
Le Chonk runs on colibri.
@MistralAI Large 4, 1T parameters, from @UnslothAI's GGUFs: text, pictures, reasoning and tools, on one 24 GB card with the experts read from NVMe. 1.34-1.36 tok/s.
colibri v2.1.0: github.com/JustVugg/colibri
Release notes and downloads:
github.com/JustVugg/colibri/…
One step to start: download, then START-HERE. Free and open source.
This is just the beginning Lumabri with Colibrì! Thank you @mfranz_on
You can chat live on 4× Raspberry Pi 5 4GB with weights split over Ethernet at 20 tok/s
No GPU, no cloud just the power of Colibri @justvugg
Local AI is never be so cool.
