π± 100% Offline & Secure Assistant on your phone
π€« No Servers. No Logins. No Tracking. No Data Uploading.
π΄ TOP SECRET INTELLIGENCE
Try it today!π
Google just dropped Gemma 4 and it's not even close to fair. @googlegemma
Gemma-4-E2B-it, An edge-to-beyond model that runs on your phone and punches way above its weight. Multimodal, 140+ languages, beats models 20x its size. This thing is absurd for a local model.
We already support it in Secret AI. Load it up, no account needed, fully offline, and your data never leaves your device.
Why wait for corporate APIs when you can run it on your own hardware?
LFM2.5-350M is out, and itβs actually pretty good. @liquidai
Tried it in Secret AI today and came away impressed. Fast (100+ tok/s), lightweight, and feels like a really nice fit for local use πππ
If roleplay isnβt your thing, hereβs Qwen3.5-2B on iPhone in Assistant Mode.
You get the familiar chatbot experience, plus customizable features like chat backgrounds, avatars, and system prompts for a more personalized everyday conversation.
Qwen 3.5 by @Alibaba_Qwen is wild β surprisingly strong for its size.
We tested it with our new feature "Characters" in Secret AI, and my iPhone turns into a private roleplay engine!
Hereβs what Qwen3.5-2B looks like running on my phone. Fully local, fast, no cloud, no account. This is what on-device AI should feel like.
The "Right to Compute Act" guarantees that individuals and organizations can own and use computational resources β hardware, software, algorithms, even quantum systems β unless the government can show a compelling reason to restrict them. It pairs that freedom with sensible guardrails for critical infrastructure, requiring companies to follow national safety frameworks like NISTβs AI Risk Management Framework. @DispatchAlerts@taylordbarkley
@linchi360@liquidai Yeah, weβd love to provide an API β itβs actively in our dev roadmap, though we donβt have an exact release date yet.
Btw, is there any pain point with Ollama thatβs stopping you from using it to achieve this? Would love to hear your thoughts
Itβs just a quick demo β extracting math formulas from images and analyzing them is still pretty heavy for a phone, and we canβt yet run the strongest models locally. But youβre right β weβll try different models later to show what kinds of problems they can actually handle.
As for your question about internet access (sorry for missing your DM!), Iβm not totally sure what you mean β are you asking if we could provide an API for it?
Meet our newest nano model: LFM2-ColBERT-350M βοΈ
At only 350M parameters, LFM2-ColBERT-350M allows you to store documents in one language and retrieve them in many languages with high accuracy and inference speeds of models a fraction of its size.
> Best cross-lingual retriever in the sub-500M class
> Outperforms larger models in German, Arabic, Korean, Spanish, Portuguese, Italian, French and Japanese
> Performs on par with much larger models in English
> Compact 350M design ready for large-scale and on-device retrieval
> Scales linearly with batch size, sustaining over 1K docs/sec in document encoding
1/n π§΅
Big news: Introducing Qwen3-Max-Preview (Instruct) β our biggest model yet, with over 1 trillion parameters! π
Now available via Qwen Chat & Alibaba Cloud API.
Benchmarks show it beats our previous best, Qwen3-235B-A22B-2507. Internal tests + early user feedback confirm: stronger performance, broader knowledge, better at conversations, agentic tasks & instruction following.
Scaling works β and the official release will surprise you even more. Stay tuned!
Qwen Chat: https://t.co/V7RmqMaVNZ
Alibaba Cloud API: https://t.co/zjCKdWee5v
@ggerganov Totally agree! These models are making local AI so much better.
And big thanks to the llama.cpp community - you've made it possible to actually run these models on consumer hardware - even on the phone! This is amazingπ
π₯ MiniCPM-V 4.5 is now LIVE on Secret AI!
Experience cutting-edge vision AI that processes images 100% offline on your device. No internet needed, no data tracking, complete privacy!
Huge thanks to @OpenBMB for this incredible breakthrough! π