OpenWeights: Edge Inference
Tinejdžeri
info
1+
Preuzimanja
O aplikaciji
OpenWeights runs open-weight language models directly on your phone. There is no account to create, no server to talk to, and nothing measured about you.
Search Hugging Face from inside the app, find out whether a model will actually run on your device before you download it, and chat with it. Every token is produced by your own hardware.
ANY MODEL, NOT A CATALOGUE
Other on-device apps hand you a short list somebody else chose. OpenWeights hands you the Hub. If someone published a GGUF an hour ago, you can run it.
HONEST ABOUT YOUR DEVICE
Before you spend gigabytes, the app reads the model's header over the network and tells you what it needs at the context length you picked, roughly how fast it will be, and whether it will run at all on your phone.
REAL NUMBERS, IN FRONT OF YOU
Tokens per second, time to first token, and how full the context window is, shown while you chat rather than hidden.
YOURS TO TUNE
Temperature, top-p, top-k, repeat penalty, context length, the system prompt and what the model is told about its tools, all saved per model. Where a model can think before it answers that is a switch too, and on a phone with a working GPU you can say which processor holds the layers.
MORE THAN TEXT
Images and audio for models that ship a projector for them, documents any model can read, dictation through your phone's own on-device recogniser, and read-aloud.
AN ASSISTANT THAT CAN ACT
The model can search the web, read a page, and work with a folder you choose to share. You decide how much rope it gets: approve every step, let it run, or ask it to plan first and say what it would do before anything happens. It can ask you a clarifying question and follow a checklist you can tick yourself. Every tool can be switched off.
PRIVATE BY CONSTRUCTION
Your conversations, your models and your usage stay on the device. There is no analytics SDK, no crash reporter, and no backend of ours. Automatic backup and device transfer are switched off, so a new phone does not inherit your chats. Uninstalling removes everything.
Two tools do reach the internet, because they have to: web search and page fetching. Both are listed under a heading that says they leave the device, both can be switched off, and every call one makes is a row in the reply that names it and what it was given. Your Hugging Face token, if you set one, is encrypted with a key held in the Android Keystore and is sent only to Hugging Face.
WHAT YOU NEED
Android 12 or newer, a 64-bit ARM device, and enough memory for the model you choose. The app will tell you before you download.
Models come from third parties. You choose which one to run, and what it says is its publisher's work rather than ours. Open source, at github.com/alpharomercoma/openweights.
Search Hugging Face from inside the app, find out whether a model will actually run on your device before you download it, and chat with it. Every token is produced by your own hardware.
ANY MODEL, NOT A CATALOGUE
Other on-device apps hand you a short list somebody else chose. OpenWeights hands you the Hub. If someone published a GGUF an hour ago, you can run it.
HONEST ABOUT YOUR DEVICE
Before you spend gigabytes, the app reads the model's header over the network and tells you what it needs at the context length you picked, roughly how fast it will be, and whether it will run at all on your phone.
REAL NUMBERS, IN FRONT OF YOU
Tokens per second, time to first token, and how full the context window is, shown while you chat rather than hidden.
YOURS TO TUNE
Temperature, top-p, top-k, repeat penalty, context length, the system prompt and what the model is told about its tools, all saved per model. Where a model can think before it answers that is a switch too, and on a phone with a working GPU you can say which processor holds the layers.
MORE THAN TEXT
Images and audio for models that ship a projector for them, documents any model can read, dictation through your phone's own on-device recogniser, and read-aloud.
AN ASSISTANT THAT CAN ACT
The model can search the web, read a page, and work with a folder you choose to share. You decide how much rope it gets: approve every step, let it run, or ask it to plan first and say what it would do before anything happens. It can ask you a clarifying question and follow a checklist you can tick yourself. Every tool can be switched off.
PRIVATE BY CONSTRUCTION
Your conversations, your models and your usage stay on the device. There is no analytics SDK, no crash reporter, and no backend of ours. Automatic backup and device transfer are switched off, so a new phone does not inherit your chats. Uninstalling removes everything.
Two tools do reach the internet, because they have to: web search and page fetching. Both are listed under a heading that says they leave the device, both can be switched off, and every call one makes is a row in the reply that names it and what it was given. Your Hugging Face token, if you set one, is encrypted with a key held in the Android Keystore and is sent only to Hugging Face.
WHAT YOU NEED
Android 12 or newer, a 64-bit ARM device, and enough memory for the model you choose. The app will tell you before you download.
Models come from third parties. You choose which one to run, and what it says is its publisher's work rather than ours. Open source, at github.com/alpharomercoma/openweights.
Run open-weight AI models on your phone. No account, no cloud, no telemetry.
Ažurirano dana
Sigurnost počinje razumijevanjem na koji način programeri prikupljaju i dijele vaše podatke. Privatnost podataka i sigurnosne prakse se mogu razlikovati ovisno o korištenju, regiji i dobi. Programer je naveo ove informacije i može ih s vremenom ažurirati.
- Podaci se ne dijele s trećim stranamaSaznajte više o načinu na koji programeri pružaju izjavu o dijeljenju
- Podaci se ne prikupljajuSaznajte više o načinu na koji programeri pružaju izjavu o prikupljanju
Šta je novo
Image detail, search width, and a breakdown of every reply
- Added a per-reply breakdown: long-press a reply for its timings, what the cache reused, and prefill and decode speeds.
- Added an image detail setting for vision models, trading speed against how much of a picture is read.
- Added a setting for how many results a search returns.
- Added a cancel button to downloads, and copying a reply as Markdown.
- Improved web page reading and pasted links.
- Fixed a greeting answered as a date.
- Added a per-reply breakdown: long-press a reply for its timings, what the cache reused, and prefill and decode speeds.
- Added an image detail setting for vision models, trading speed against how much of a picture is read.
- Added a setting for how many results a search returns.
- Added a cancel button to downloads, and copying a reply as Markdown.
- Improved web page reading and pasted links.
- Fixed a greeting answered as a date.
Podrška za aplikaciju
Informacije o programeru
Alpha Romer N. Coma
alpha.coma.ict@gmail.com
Blk 193 LOT 16 17th Street Phase 5B
Meycauayan
3020
Philippines
