It would be great if the local llm have access to local tools you can enable/disable as needed (e.g. via customizable profiles). Simple tools such as fetch url, file access, messaging, calendar, etc would be very useful, though I'm not sure if the input token limit is large enough to allow this. Even better if it can somehow do web search but I understand it would be hard to do for free.
Also, how cool it would be if you can expose openai compatible api that can be accessed from other devices in your local network? Imagine turning your old phones into local llm servers. That would be very cool.
By the way, I can't figure out how to clear previous chats data. Is it hidden somewhere?
> However, both are platform-specific and only support specific models from the company
This is not true, as you are for sure aware. Google AI edge supports a lot models, including any Litert model from huggingface, pytorch ones etc. [0]. Additionally, it's not even platform specific, works for iOS [1].
Why lie? I understand that your framework does more stuff like MCP, but I'm sure that's coming for Google's as well. I guess if the UX is really better it can work, but i would also say Ollama's use cases are quite different because on desktop there's a big community of hobbyists that cook up their own little pipelines/just chat to LLMs with local models (apart from the desktop app devs). But on phones, imo that segment is much smaller. App devs are more likely to use the 1st party frameworks, rather than 3rd party. I wouldnt even be surprised if apple locks down at some points some API's for safety/security reasons.
Thanks for the feedback. You're right to point out that Google AI Edge is cross-platform and more flexible than our phrasing suggested.
The core distinction is in the ecosystem: Google AI Edge runs tflite models, whereas Cactus is built for GGUF. This is a critical difference for developers who want to use the latest open-source models.
One major outcome of this is model availability. New open source models are released in GGUF format almost immediately. Finding or reliably converting them to tflite is often a pain. With Cactus, you can run new GGUF models on the day they drop on Huggingface.
Quantization level also plays a role. GGUF has mature support for quantization far below 8-bit. This is effectively essential for mobile. Sub-8-bit support in TFLite is still highly experimental and not broadly applicable.
Last, Cactus excels at CPU inference. While tflite is great, its peak performance often relies on specific hardware accelerators (GPUs, DSPs). GGUF is designed for exceptional performance on standard CPUs, offering a more consistent baseline across the wide variety of devices that app developers have to support.
GGUF is more suitable for the latest open-source models, i agree there. Quant2/Q4 will probably be critical as well, if we don't see a jump in ram. But then again I wonder when/If mediapipe will support GGUF as well.
PS, I see you are in the latest YC batch? (below you mentioned BF). Good luck and have fun!
I would say that while Google's MediaPipe can technically run any tflite model, it turned out to be a lot more difficult to do in practice with third-party models compared to the "officially supported" models like Gemma-3n. I was trying to set up a VLM inference pipeline using a SmolVLM model. Even after converting it to a tfilte-compatible binary, I struggled to get it working and then once it did work, it was super slow and was obviously missing some hardware acceleration.
I have not looked at OP's work yet, but if it makes the task easier, I would opt for that instead of Google's "MediaPipe" API.
Is this using only llama.cpp as inference engine? How is this days support there on NPU and GPU? Not sure if LLM can run on NPU but many models like STT and TTS and vision often can run much faster on Apple NPU
Probably they mean new apps. Since kotlin multiplatform on android is just native android and android share is like 70% if devices it already at least 50% market share of mobile apps. If you add flutter and react native there is not much left: only games like unity and unreal. I see much less iOS jobs these days.
Ollama runs on Android just fine via Termux. I use it with 5GB models. They even recently added ollama package, there is no longer need to compile it from source code.
How does it work? How does one model on the device get shared to many apps? Does each app have it's own inference sdk running or is there one inference engine shared to many apps (like ollama does). If it's the later, what's the communication protocol to the inference engine?
Great question. Currently, each app is sandboxed - so each model file is downloaded inside each app's sandbox. We're working on enabling file sharing across multiple apps so you don't have to redownload the model.
With respect to the inference SDK, yes you'll need to install the (react native/flutter) framework inside each app you're building.
The SDK is very lightweight (our own iOS app is <30MB which includes the inference SDK and a ton of other stuff)
I would like to see it as an app, tbh! If I could run it as an APK with a nice GUI interface for picking different models to run, that would be a killer feature.
Thank you especially for the phone model vs tok/s breakdown. Do you have such tables for more models? For models even leaner than Gemma3 1B. How low can you go? Say if I wanted to tweak out 45toks/s on an iPhone 13?
P.S: Also, I'm assuming the speeds stay consistent with react-native vs. flutter etc?
thank you! We're continue to add performance metrics as more data comes in.
A Qwen 2.5 500M will get you to ≈45tok/sec on an iPhone 13. Inference speeds are somewhat linearly inversely proportional to model sizes.
Yes, speeds are consistent across frameworks, although (and don't quote me on this), I believe React Native is slightly slower because it interfaces with the C++ engine through a set of bridges.
I also want to add on that I really appreciate the benchmarks.
When I was working with RAG llama.cpp through RN early last year I had pretty acceptable tok/sec results up through 7-8b quantized models (on phones like the S24+ and iPhone 15pro). MLC was definitely higher tok/sec but it is really tough to beat the community support and availability in the gguf ecosystem.
Looking at the current benchmarks table, I was curious: what do you think is wrong with Samsung S25 Ultra?
Most of the standard mobile CPU benchmarks (GeekBench, AnTuTu, et al) show a 20-40% performance gain over S23/S24 Ultra. Also, this bucks the trend where most other devices are ranked appropriately (i.e. newer devices perform better).
appreciate the feedback! Made the demo links more prominent on the README.
Some Android models won't support GPU hardware; we'll be addressing that as we move to our own kernels.
The app itself is just a demonstration of Cactus performance. The underlying framework gives you the tools to build any local mobile AI experience you'd like.
Thank you! it goes without saying that the field is rapidly developing, so the use cases range from private AI assistant/companion apps to internet connectivity-independent copilots to powering private wearables, etc.
We're currently working with a few projects in the space.
For argument's sake, suppose we live in a world where many high-quality models can be run on-device. Is there any concern from companies/model developers about exposing their proprietary weights to the end user? It's generally not difficult to intercept traffic (weights) sent to and app, or just reverse the app itself.
Simple answer is they won't send the model to the end user if they don't want it used outside their app.
This isn't really anything novel to LLMs of AI models. Part of the reason for many previously desktop applications being cloud or requiring cloud access is keeping their sensitive IP off the end users' device.
So far, our focus is on supporting models with fully open-sourced weights. Providers who are sensitive about their weights typically lock those weights up in their cloud and don't run their models locally on consumer devices anyway.
I believe there are some frameworks pioneering model encryption, but i think we're a few steps away from wide adoption.
GGUF is easy to implement, but you'd probably find better performance with tflite on mobile for their custom XNNPACK kernels. Performance is pretty critical on low-power devices.
We are writing our own backend, but tflite (now called LiteRT) was not faster than GGML when we tested and GGML is already well supported. But we are moving away completely anyway.
What do you think about security? I mean, a model with full (or partial) access to the smartphone and internet. Even if it runs locally, isn't there still a risk that these models could gain full access to the internet and the device?
The models themselves live in an isolated sandbox. On top of that, each mobile app has its own sandbox - isolated from the phone's data or tools.
Both the model and the app only have access to the tools or data that you choose to give it. If you choose to give the model access to web search - sure, it'll have (read-only) access to internet data.
Idk who these people are and I am sure they have good intentions, but they're wrapping llama.cpp.
That's what "like Ollama" means when you're writing code. That's also why there's a ton of comments asking if it's a server or app or what (it's a framework that an app would be built to use, you can't have an app with a localhost server like ollama on Android & iOS)
There's plenty of projects much further ahead, and I don't appreciate the amount of times I've seen this project come up in conversation the past 24 hours, due to misleading assertions that looked LLM-written, and a rush to make marketing claims that are just stuff llama.cpp does for you.
Flutter isolates are like threads, and servers may use multithreading to handle requests, and Ollama is like a server in that it provides an API, and since we've shown both are servers, it's like Ollama?
Please do educate me on this, I'm fascinated.
When you're done there, let's say Flutter having isolates does mean you have a React Native and Flutter local LLM server.
What's your plan for your Android & iOS-only framework being a system server? Or alternatively, available at localhost for all apps to contact?
We are following Ollama's design, but not verbatim due to apps being sandboxed.
Phones are resource-constrained, we saw significant battery overhead with in-process HTTP listeners so we stuck with simple stateful isolates in Flutter and exploring standalone server app others can talk to for React.
For model sharing with the current setup:
iOS - We are working towards writing the model into an App Group container, tricky but working around it.
Android - We are working towards prompting the user once for a SAF directory (e.g., /Download/llm_models), save the model there, then publish a ContentProvider URI for zero-copy reads.
We are already writing more mobile-friendly kernels and Tensors, but GGML/GGUF is widely supported, porting it is an easy way to get started and collect feedback, but we will completely move away from in < 2 months.
How does writing a model into an App Group container enable your framework to enable an app to enable a local LLM server that 3rd party apps can make calls to on iOS?[^1]
How does writing a model into a shared directory on Android enable a local LLM server that 3rd party apps can make calls to?[^2]
How does writing your own kernels get you off GGUF in 2 months? GGUF is a storage format. You use kernels to do things with the numbers you get from it.
I thought GGUF was an advantage? Now it's something you're basically done using?
I don't think you should continue this conversation. As easy it as it is to get your work out there, it's just as easy to build a record of stretching truth over and over again.
Best of luck, and I mean it. Just, memento mori: be honest and humble along the way. This is something you will look back on in a year and grimace.
[^1] App group containers only work between apps signed from the same Apple developer account. Additionally, that is shared storage, not a way to provide APIs to other apps.
[^2] SAF = Storage Access Framework, that is shared storage, not a way to provide APIs to other apps.
I was merely replying to questions, thinking you were genuinely asking. On reviewing this thread, I feel you are angry for some reason and taking my responses out of context. I clearly explained that running a server came with battery overhead, so we ditched it in favour of shared app group for model weight sharing. Anyway, Thanks and have a nice day
The best way to go about this is realizing that there are more people reading this thread that make their own assumptions.
Not staying professional and just answering the questions, and just doing "aight im outta here" when it gets a little bit harder is not a good look; it seems like you can't defend your own project.
- "You are, undoubtedly, the worst pirate i have ever heard of"
- "Ah, but you have heard of me"
Yes, we are indeed a young project. Not two weeks, but a couple of months. Welcome to AI, most projects are young :)
Yes, we are wrapping llama.cpp. For now. Ollama too began wrapping llama.cpp. That is the mission of open-source software - to enable the community to build on each others' progress.
We're enabling the first cross-platform in-app inference experience for GGUF models and we're soon shipping our own inference kernels fully optimized for mobile to speed up the performance. Stay tuned.
Does this download models at runtime? I would have expected a different API for that. I understand that you don’t want to include a multi-gig model in your app. But the mobile flow is usually to block functionality with a progress bar on first run. Downloading inline doesn’t integrate well into that.
You’d want an API for downloading OR pulling from a cache. Return an identifier from that and plug it into the inference API.
It would be great if the local llm have access to local tools you can enable/disable as needed (e.g. via customizable profiles). Simple tools such as fetch url, file access, messaging, calendar, etc would be very useful, though I'm not sure if the input token limit is large enough to allow this. Even better if it can somehow do web search but I understand it would be hard to do for free.
Also, how cool it would be if you can expose openai compatible api that can be accessed from other devices in your local network? Imagine turning your old phones into local llm servers. That would be very cool.
By the way, I can't figure out how to clear previous chats data. Is it hidden somewhere?
This is not true, as you are for sure aware. Google AI edge supports a lot models, including any Litert model from huggingface, pytorch ones etc. [0]. Additionally, it's not even platform specific, works for iOS [1].
Why lie? I understand that your framework does more stuff like MCP, but I'm sure that's coming for Google's as well. I guess if the UX is really better it can work, but i would also say Ollama's use cases are quite different because on desktop there's a big community of hobbyists that cook up their own little pipelines/just chat to LLMs with local models (apart from the desktop app devs). But on phones, imo that segment is much smaller. App devs are more likely to use the 1st party frameworks, rather than 3rd party. I wouldnt even be surprised if apple locks down at some points some API's for safety/security reasons.
[0] https://ai.google.dev/edge/mediapipe/solutions/genai/llm_inf...
[1] https://ai.google.dev/edge/mediapipe/solutions/genai/llm_inf...
The core distinction is in the ecosystem: Google AI Edge runs tflite models, whereas Cactus is built for GGUF. This is a critical difference for developers who want to use the latest open-source models.
One major outcome of this is model availability. New open source models are released in GGUF format almost immediately. Finding or reliably converting them to tflite is often a pain. With Cactus, you can run new GGUF models on the day they drop on Huggingface.
Quantization level also plays a role. GGUF has mature support for quantization far below 8-bit. This is effectively essential for mobile. Sub-8-bit support in TFLite is still highly experimental and not broadly applicable.
Last, Cactus excels at CPU inference. While tflite is great, its peak performance often relies on specific hardware accelerators (GPUs, DSPs). GGUF is designed for exceptional performance on standard CPUs, offering a more consistent baseline across the wide variety of devices that app developers have to support.
GGUF is more suitable for the latest open-source models, i agree there. Quant2/Q4 will probably be critical as well, if we don't see a jump in ram. But then again I wonder when/If mediapipe will support GGUF as well.
PS, I see you are in the latest YC batch? (below you mentioned BF). Good luck and have fun!
I have not looked at OP's work yet, but if it makes the task easier, I would opt for that instead of Google's "MediaPipe" API.
Is this really true? Where are these stats coming from?
We are a dev toolkit to run LLMs cross-platform locally in any app you like.
With respect to the inference SDK, yes you'll need to install the (react native/flutter) framework inside each app you're building.
The SDK is very lightweight (our own iOS app is <30MB which includes the inference SDK and a ton of other stuff)
We are working on agentic browser (also launched today https://news.ycombinator.com/item?id=44523409 :))
Right now we have a desktop version with ollama support, but we want to build a mobile chromium fork with local LLM support. Will check out cactus!
DM me on BF - let's talk!
Would be great to have a few larger models to choose from too, Qwen 3 4b, 8b etc
Thank you especially for the phone model vs tok/s breakdown. Do you have such tables for more models? For models even leaner than Gemma3 1B. How low can you go? Say if I wanted to tweak out 45toks/s on an iPhone 13?
P.S: Also, I'm assuming the speeds stay consistent with react-native vs. flutter etc?
A Qwen 2.5 500M will get you to ≈45tok/sec on an iPhone 13. Inference speeds are somewhat linearly inversely proportional to model sizes.
Yes, speeds are consistent across frameworks, although (and don't quote me on this), I believe React Native is slightly slower because it interfaces with the C++ engine through a set of bridges.
When I was working with RAG llama.cpp through RN early last year I had pretty acceptable tok/sec results up through 7-8b quantized models (on phones like the S24+ and iPhone 15pro). MLC was definitely higher tok/sec but it is really tough to beat the community support and availability in the gguf ecosystem.
Most of the standard mobile CPU benchmarks (GeekBench, AnTuTu, et al) show a 20-40% performance gain over S23/S24 Ultra. Also, this bucks the trend where most other devices are ranked appropriately (i.e. newer devices perform better).
Thanks for sharing your project.
S25 is an outlier that surprised us too.
I got $10 on S25 climbing back up to the top of the rankings as more data comes in :)
The performance is quite good, even on CPU.
However I'm now trying it on a pixel, and it's not using GPU if I enable it.
I do like this idea as I've been running models in termux until now.
Is the plan to make this app something similar to lmstudio for phones?
Some Android models won't support GPU hardware; we'll be addressing that as we move to our own kernels.
The app itself is just a demonstration of Cactus performance. The underlying framework gives you the tools to build any local mobile AI experience you'd like.
can you tell us more about the use cases that you have in mind? I saw that you're able to run 1-4B models (which is impressive!)
We're currently working with a few projects in the space.
For a demo of a familiar chat interface, download https://apps.apple.com/gb/app/cactus-chat/id6744444212 or https://play.google.com/store/apps/details?id=com.rshemetsub...
For other applications, join the discord and stay tuned! :)
This isn't really anything novel to LLMs of AI models. Part of the reason for many previously desktop applications being cloud or requiring cloud access is keeping their sensitive IP off the end users' device.
I believe there are some frameworks pioneering model encryption, but i think we're a few steps away from wide adoption.
Both the model and the app only have access to the tools or data that you choose to give it. If you choose to give the model access to web search - sure, it'll have (read-only) access to internet data.
Idk who these people are and I am sure they have good intentions, but they're wrapping llama.cpp.
That's what "like Ollama" means when you're writing code. That's also why there's a ton of comments asking if it's a server or app or what (it's a framework that an app would be built to use, you can't have an app with a localhost server like ollama on Android & iOS)
There's plenty of projects much further ahead, and I don't appreciate the amount of times I've seen this project come up in conversation the past 24 hours, due to misleading assertions that looked LLM-written, and a rush to make marketing claims that are just stuff llama.cpp does for you.
1) The commit history goes back to April.
2) LlaMa.cpp licence is included in the Repo where necessary like Ollama, until it is deprecated.
3) Flutter isolates behave like servers, and Cactus codes use that.
Flutter isolates are like threads, and servers may use multithreading to handle requests, and Ollama is like a server in that it provides an API, and since we've shown both are servers, it's like Ollama?
Please do educate me on this, I'm fascinated.
When you're done there, let's say Flutter having isolates does mean you have a React Native and Flutter local LLM server.
What's your plan for your Android & iOS-only framework being a system server? Or alternatively, available at localhost for all apps to contact?
Phones are resource-constrained, we saw significant battery overhead with in-process HTTP listeners so we stuck with simple stateful isolates in Flutter and exploring standalone server app others can talk to for React.
For model sharing with the current setup:
iOS - We are working towards writing the model into an App Group container, tricky but working around it.
Android - We are working towards prompting the user once for a SAF directory (e.g., /Download/llm_models), save the model there, then publish a ContentProvider URI for zero-copy reads.
We are already writing more mobile-friendly kernels and Tensors, but GGML/GGUF is widely supported, porting it is an easy way to get started and collect feedback, but we will completely move away from in < 2 months.
Anything else you would like to know?
How does writing a model into a shared directory on Android enable a local LLM server that 3rd party apps can make calls to?[^2]
How does writing your own kernels get you off GGUF in 2 months? GGUF is a storage format. You use kernels to do things with the numbers you get from it.
I thought GGUF was an advantage? Now it's something you're basically done using?
I don't think you should continue this conversation. As easy it as it is to get your work out there, it's just as easy to build a record of stretching truth over and over again.
Best of luck, and I mean it. Just, memento mori: be honest and humble along the way. This is something you will look back on in a year and grimace.
[^1] App group containers only work between apps signed from the same Apple developer account. Additionally, that is shared storage, not a way to provide APIs to other apps.
[^2] SAF = Storage Access Framework, that is shared storage, not a way to provide APIs to other apps.
Not staying professional and just answering the questions, and just doing "aight im outta here" when it gets a little bit harder is not a good look; it seems like you can't defend your own project.
Just FYI.
Good luck!
- "You are, undoubtedly, the worst pirate i have ever heard of" - "Ah, but you have heard of me"
Yes, we are indeed a young project. Not two weeks, but a couple of months. Welcome to AI, most projects are young :)
Yes, we are wrapping llama.cpp. For now. Ollama too began wrapping llama.cpp. That is the mission of open-source software - to enable the community to build on each others' progress.
We're enabling the first cross-platform in-app inference experience for GGUF models and we're soon shipping our own inference kernels fully optimized for mobile to speed up the performance. Stay tuned.
PS - we're up to good (source: trust us)
we support cloud fallback as an add-on feature. This lets us support vision and audio in addition to text.
Most projects typically start with llama.cpp and then move away to proprietary kernels
https://github.com/cactus-compute/cactus/tree/main/react#emb...
(Flutter works the same way)
What are you building?
You’d want an API for downloading OR pulling from a cache. Return an identifier from that and plug it into the inference API.
We're restructuring the model initialization API to point to a local file & exposing a separate abstracted download function that takes in a URL.
wrt downloading post-install: based on our feedback, this is indeed a preferred pattern (as opposed to bundling in large files).
We'll update the download API, thanks again.
https://play.google.com/store/apps/details?id=com.rshemetsub...
Otherwise, it's easy to build any of the example apps from the repo:
cd react/example && yarn && npx expo run:android
or
cd flutter/example && flutter pub get && flutter run