Spaces:
Running
[NEW] HuggingChat Omni
Introducing: HuggingChat Omni π«
HuggingChat returns and it's smarter and faster than ever π
Stop picking models. Start chatting.
- 115+ available models - https://huggingface.co/chat/models
- 15+ providers available - powered by Hugging Face Inference Providers.
- One chat interface: HuggingChat
Available now for all Hugging Face users. Free users can use their inference credits, PRO users get 20x more credits to use.
π§ Omni: the new default routing model
When you send a message, Omni analyzes what you need and routes you to the best model for that specific task.
Each route uses the best model for its task. You see which model handled your request while it streams.
π Examples
| What you ask | Route | Model |
|---|---|---|
| "Help me decide between two job offers. One pays 20% more but requires relocation." | decision_support |
deepseek-ai/DeepSeek-R1-0528 |
| "Create a React component for an image carousel with lazy loading" | code_generation |
Qwen/Qwen3-Coder-480B-A35B-Instruct |
| "Write a short mystery story set in a lighthouse during a storm" | creative_writing |
moonshotai/Kimi-K2-Instruct-0905 |
| "Translate this to French: The meeting has been rescheduled to next Tuesday" | translation |
CohereLabs/command-a-translate-08-2025 |
βοΈ Under the hood
Omni uses a policy-based routing system. Each route has:
- A clear description of what it handles
- A primary model best suited for that task
- Fallback models if the primary is unavailable
The router model analyzes your conversation and picks the matching route. Fast (10 second timeout) and runs on every message. Credits to Katanemo for their routing model: katanemo/Arch-Router-1.5B
β¨ What else is new
- Background generation tracking: Multiple conversations can generate at the same time. Switch between tabs and the app tracks what's still generating. Updates appear automatically when responses finish.
- Better streaming: Text renders faster and smoother. The app only updates what changed instead of re-rendering everything. Less flickering, especially in long responses with code blocks.
- Better UX: UX was refined throughout the app. Fewer bugs and rough edges. Preview for code, beautiful streaming and more polish and attention to detail everywhere.
- Speed optimizations: Sessions stay active longer with automatic token refresh. Response times improved across the board. The whole app feels faster.
π οΈ Run it yourself
HuggingChat is of course still 100% open source. It has never been easier to self-host your own instance.
Quick setup:
git clone https://github.com/huggingface/chat-ui
cd chat-ui
npm install
npm run dev
Only 3 env variables to set to get it working in .env:
MONGODB_URL- Your MongoDB connectionOPENAI_API_KEY- Your API keyOPENAI_BASE_URL- Your endpoint URL
You can also configure your own routes in a JSON file. Each route defines which models to use for specific tasks.
Check out the repo: github.com/huggingface/chat-ui
Hope you are as excited as we are about HuggingChat Omni! Please share your feedback and ideas in this thread π€
Is it possible to import my conversations from the previous version of HuggingChat?
Yeah this dumbing down the system was totally worth nuking everyone's logs and assistants...? The performance improvements are nice if true, but how can you call this a better UX when so many basic features are missing from the last version? Even simple settings are gone, like no options to delete or edit output? There isn't even a way to tweak temperature/repetition minimizing settings, or give different chats different system prompts??
wow, I'm kind of surprised it's back. feels like a tad bit of a downgrade, but I'm assuming that it was a complete rework? hoping that more QoL features will be reintroduced again.
Does the platform now impose usage limits based on inference credits for free users?
we're so back
edit:
nevermind, cant delete the conversation branch like beforeπ’
edit 2:
and it now has a limit. Its been over six hours and i still cant continue the conversation π
Thanks for getting this running
Import please? that's the only reason why I still return to this...
Please add back the message deletion feature, sometimes it feels gets too much messages and gets lost easily
Making advertising for another ai on the official hugging chat forum is kinda dirty if you ask me
I'm kinda dumb, but where are my chats? Are they like gone? I didn't see a option to export it while HC was down.
Yes, your chats are gone. There was an option to export for a couple weeks after chat went down, then they were deleted.
Thanks for your feedback everyone <3
Is it possible to import my conversations from the previous version of HuggingChat?
Not really but we can work on it if a lot of people want it.
Can't edit conversations
Mhhh you can (or I'm missing something)
Can't change temperature
Can't search the web
Can't generate imagesWe plan to add this for the rest I'm not sure yet but Assistants may come back (and they will be better)
β You now have to pay to use it π
There's a free tier. It's way faster and stable now also PRO Subscription is a very nice deal https://huggingface.co/pro. Also paid requests are very cheap for most models. Example from my https://huggingface.co/settings/inference-providers/overview: 1840 requests with
Qwen/Qwen3-235B-A22B-Instruct-2507= .65$tried it out a lot, seems pretty much good as new. new UI looks different for sure. works well, haven't had it crash at all and haven't seen any of the weird rendering bugs that were so prevalent in the old version.
Thanks that's motivating and yes we'll make it awesome with faster iteration rate than the 1st version!
Imagine using it like the old HuggingChat. You never can. You do have to pay to use it now becasuse it is no longer unlimited like it used to be. For unlimited messages, this is not a good chat interface. I thought this was about open-source freedom, not the money.
The "CohereLabs-AyaExpanse" model seems to work well for my purposes, given i just porked Calamity Coyote, with no "nudging" or other tricks needed. Still doesn't solve the problem of limited credits per months, and whatever, but at least it's something, for now. Otherwise there's venice.ai which is completely uncensored, but also much lower quality.