Hey HN, Henry & Roman here, we are building Cactus (https://cactuscompute.com/), an AI inference engine specifically designed for phones.

We're seeing a major push towards on-device AI, and for good reason: on-device AI decreases latency from >1sec to <100ms, guarantees privacy by default, works offline, and doesn't rack up a massive API bill at scale.

Also, tools and agentic designs make small models really good beyond benchmarks. This has been corroborated by other papers like https://arxiv.org/abs/2506.02153, and we see model companies like DeepMind aggressively going into smaller models with Gemma3 270m and 308m. We found Qwen3 600m to be great at tool calls for instance.

Some frameworks already try to solve this but in my previous job, they struggled in production compared to research and playgrounds:

- They optimise for modern devices but 70% of phones today are low-mid budget.

- Bloated app bundle sizes and battery drain are serious concerns for users.

- Phone GPU battery drain is unacceptable, NPUs are preferred, but few phones have those for now.

- Some are platform-specific, requiring different models and workflows for different operating systems.

At Cactus, we’ve written kernels and inference engine for running AI locally on any phone, from the ground-up.

Cactus is designed for mobile devices and their constraints. Every design choice like energy efficiency, accelerator support, quantization levels, supported models, weight format, and context management were determined by this. We also provide minimalist SDKs for app developers to build agentic workflows in 2-5 lines of code.

We made a Show HN post when we started the project to get the community's thoughts (https://news.ycombinator.com/item?id=44524544). Based on your feedback, we built Cactus bottom-up to solve those problems, and are launching the Cactus Kernels, Cactus Graph and Cactus Engine, all designed for phones and tiny devices.

CPU benchmarks for Qwen3-600m-INT8 :

- 16-20 toks/sec on Pixel 6a / Galaxy S21 / iPhone 11 Pro

- 50-70 toks/sec on Pixel 9 / Galaxy S25 / iPhone 16.

- Time-to-first-token is as low as 50ms depending on prompt size.

On NPUs, we see Qwen3-4B-INT4 run at 21 toks/sec.

We are open-source (https://github.com/cactus-compute/cactus). Cactus is free for hobbyists and personal projects, with a paid license required for commercial use.

We have a demo app on the App Store at https://apps.apple.com/gb/app/cactus-chat/id6744444212 and on Google Play at https://play.google.com/store/apps/details?id=com.rshemetsub....

In addition, there are numerous apps using Cactus in production, including AnythingLLM (https://anythingllm.com/mobile) and KinAI (https://mykin.ai/). Collectively they run over 500k weekly inference tasks in production.

While Cactus can be used for all Apple devices including Macbooks due to their design, for computers/AMD/Intel/Nvidia generally, please use HuggingFace, Llama.cpp, Ollama, vLLM, MLX. They're built for those, support x86, and are all great!

Thanks again, please share your thoughts, we’re keen to understand your views.

I've been using Cactus for a few months, great product!

Makes it really easy to plug and play different models on my phone.

If anybody is curious what a Pixel 9 Pro is capable of:

Tokens: 277- TTFT: 1609ms 9 tok/sec

qwen2.5 1.5b instruct q6_k

Sure, here's a simple implementation of the Bubble Sort algorithm in Python:

def bubble_sort(arr): n = len(arr) for i in range(n): # Flag to detect any swap in current pass swapped = False for j in range(0, n-i-1): # Swap if the element found is greater than the next element if arr[j] > arr[j+1]: arr[j], arr[j+1] = arr[j+1], arr[j] swapped = True # If no swap occurs in the inner loop, the array is already sorted if not swapped: break

# Example usage: arr = [64, 34, 25, 12, 22, 11, 90] bubble_sort(arr) print("Sorted array is:", arr)

This function sorts the array in ascending order using the Butbble Sort algorithm. The outer loop runs n times, where n is the length of the array. The inner loop runs through the array, comparing adjacent elements and swapping them if they are in the wrong order. The swapped flag is used to detect if any elements were swapped in the current pass, which would indicate that the array is already sorted and can be exited early.

Question: can this utilize multiple forms of compute at once? Many phones have both GPUs that are capable of doing compute as well as NPUs, and that number will only increase. I'm sure it would be challenging, but that's a lot of performance to leave on the table if it can't do so already.

I am very curious what could be done with your impressive optimization on an rk3588, since it has pretty decent bits in all 3 categories, and am now seriously considering a Radxa Orion to play with this on :)

One more if you have a moment: will this be limited to text generation, or will it have audio and image capabilities as well? Would be neat to enable not only image generation, but also explore voice recognition, translation, computer vision, as well as image editing and enhancement features in mobile apps beyond what the big players daign to give us :)

This paragraph is a bit confusing:

> While Cactus can be used for all Apple devices including Macbooks due to their design, for computers/AMD/Intel/Nvidia generally, please use HuggingFace, Llama.cpp, Ollama, vLLM, MLX. They're built for those, support x86, and are all great!

It reads like you're saying for all Apple devices (which would include iOS), use these other things.(?) For iOS, are you trying to beat performance of other options? If so, it would be helpful to include comparison benchmarks.

FWIW They change license 2 weeks ago from Apache 2.0 to non commercial. Understand they need to pay the bills but lost trust with such move. Will stick with react-native-ai [0] that is extension of vercel aisdk but with also local inference on edge devices

[0] react-native-ai.dev

how many GB does an app packaged with Qwen3 600m + Cactus take up?

e.g. if I built a basic LLM chat app with Qwen3 600m + Cactus, whats the total app size?

Wow made an account just for this! I was using cactus for a paid app i soft launched recently. Does that mean that now i cannot update this dependency? What is your pricing? I do not see that anywhere. If i did not notice this and pulled the updated version, would i be liable to be sued? Also, I implemented cactus on the good faith assumption that i was going to receive the updates in your roadmap, as a proper apache project. I must admit this is quite the move guys

Can you clarify the following sentence:

> We are open-source (https://github.com/cactus-compute/cactus). Cactus is free for hobbyists and personal projects, with a paid license required for commercial use.

If it is open-source, one is free to distribute even for commercial use by definition. Which one is correct and what's your business model?

How does this startup plan to make money?

Are you saying this would be more performant than Apple’s on device LLM/inferencing?

I built apps using Flutter and this project seems to make it possible to use models directly in app instead of cloud APIs. Curious about the commercial license here. How is the trade off between pricing and performance?

Will this drain my battery

Tried the android app but model download froze. Are you using the same docker-style repositories as Ollama. Because they suck. If you do I suggest use your own s3 instead.

Does it incorporate web search tool?

Would you consider adding a mode where it could go online if the user instructed it to?

The first picture on the android app store page shows Claude Haiku as the model

AI Dungeon should contact you to make an offline mode again.

curious. what are the usecases for <100ms latency ?

Can it be fine tuned for a specific task?