THIS IS BRILLIENT!!
I wanted to thank you for your hardwork.. this is incredible work!! Ornith is now better than ever! Well Done!
I'm glad you find it useful. 🙏
10 days since i started using this model, it is incredible!!! the 1M YARN extended context over an already Beast of a model Ornith 35B have proved incredibly powerful!!! it never have ONCE failed a task, errored an api, or missed the mark!! for low RAM users i really do suggest never using the embedded MTP.. and tbh i've been finding the output without the MTP is generally more focused and giving it a creative balance with temp 0.6 gives amazing result with designing UI/UX or creative content generation! Arabic language is soo much better than base Qwen's for somereason (not yet as perfect as Gemma4 but heck if u need articulated Arabic just switch to gemma4 for that task thats the power of modern open models) i can yap and rant all day about this experience but i have never completed big projects on a local model as much as i did with this model... tackled PHP, Flutter, C#, Next,js, Vue.js, Node and React, Go, Cpp, either new codebase or git clone a repo and study/modify/add features to the code bases, UI/UX for websites and whole static websites, used it on open design, used it in zoo code, opencode, claude code, TradingAgents, it never have ever failed me once. i just have to stop here.. Edit: i even forgot to mention that mind you im using the Q4 quant not even the Q8!
That’s amazing to hear that this model is so much useful to you.
I really didn’t do much myself. Basically ANY model can be made 1M context. But the problem is mostly models beyond their context limits usually degrade in their outputs in long contexts. That is why I tested and made sure the 1M context I experimented with this model shows some level of non-degraded output. The needle in a haystack verification helps assure it is good in longer contexts than this model is intended to do.
But for future models I myself will not upload 1M context models, but other different optimisation. I will only provide commands, tools to instead extend the context limit to 1M or custom context limits through my tools, so users like you can just get the context limit they desire themselves.
The risk with unverified long context will be that they may or may not be degrading in long context. But I see it’s absolutely worth doing it while acknowledging the risks and experimentation.
You MUST thank the Ornith team themselves, as the model is only their efforts, and I only just published the readymade 1M context files in this repo, verified the NIAH tests, nothing much else.
I recently also published a abliterated/uncensored version of the same Ornith 9B/35B model on my profile: https://huggingface.co/satgeze/Ornith-1.0-35B-abliterated-SatGeZe
It is my own effort using my own tools to remove any refusal guardrails, while keeping the model’s own original capabilities absolutely unchanged. There are instructions in the model to set its context limits to 1M if you need those. Hope this helps. 🙏🖖
There are bigger models I will publish which will fit on smaller VRAM and still keep capabilities they originally have in their original giant size. Keep an eye on updates from my file.
yes, thank you for raising the limit n this wonderful model.