Cactus
May 27, 2026 · View on GitHub
A low-latency AI engine for mobile devices & wearables.
- Fast & accurate: fastest inference on ARM CPU, Cactus quants at 4-bit matches f16
- Low RAM: zero-copy memory mapping ensures 10x lower RAM use than other engines
- Multimodal: one engine for speech, vision, and language models
- Cloud fallback: automatically route requests to cloud models if needed
- Model-Agnostic: Custom PyTorch models can be exported to the Cactus runtime.
┌─────────────────┐
│ Cactus Engine │ ←── OpenAI-compatible APIs for text, speech, and vision.
└─────────────────┘
│
┌─────────────────┐
│ Cactus Graph │ ←── Zero-copy computation graph ensures 10x lower RAM
└─────────────────┘
│
┌─────────────────┐
│ Cactus Kernels │ ←── Fastest ARM SIMD kernels (Apple, Samsung, Pixel, etc)
└─────────────────┘
│
┌─────────────────┐
│ Cactus Quants │ ←── Cactus Quants at 4-bit uniform matches f16.
└─────────────────┘
│
┌─────────────────┐
│Cactus Transpiler│ ←── Transpiles custom PyTorch model to Cactus.
└─────────────────┘
Quick Demo (Mac)
- Step 1:
brew install cactus-compute/cactus/cactus - Step 2:
cactus transcribeorcactus run
Cactus Engine
#include "cactus.h"
cactus_model_t model = cactus_init(
"path/to/weight/folder",
"path to txt or dir of txts for auto-rag",
false
);
const char* messages = R"([
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "My name is Henry Ndubuaku"}
])";
const char* options = R"({
"max_tokens": 50,
"stop_sequences": ["<|im_end|>"]
})";
char response[4096];
int result = cactus_complete(
model, // model handle
messages, // JSON chat messages
response, // response buffer
sizeof(response), // buffer size
options, // generation options
nullptr, // tools JSON
nullptr, // streaming callback
nullptr, // user data
nullptr, // pcm audio buffer
0 // pcm buffer size
);
Example response from Gemma3-270m
{
"success": true, // generation succeeded
"error": null, // error details if failed
"cloud_handoff": false, // true if cloud model used
"response": "Hi there!",
"function_calls": [], // parsed tool calls
"confidence": 0.8193, // model confidence
"time_to_first_token_ms": 45.23,
"total_time_ms": 163.67,
"prefill_tps": 1621.89,
"decode_tps": 168.42,
"ram_usage_mb": 245.67,
"prefill_tokens": 28,
"decode_tokens": 50,
"total_tokens": 78
}
Cactus Graph
#include "cactus.h"
CactusGraph graph;
auto a = graph.input({2, 3}, Precision::FP16);
auto b = graph.input({3, 4}, Precision::INT8);
auto x1 = graph.matmul(a, b, false);
auto x2 = graph.transpose(x1);
auto result = graph.matmul(b, x2, true);
float a_data[6] = {1.1f, 2.3f, 3.4f, 4.2f, 5.7f, 6.8f};
float b_data[12] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12};
graph.set_input(a, a_data, Precision::FP16);
graph.set_input(b, b_data, Precision::INT8);
graph.execute();
void* output_data = graph.get_output(result);
graph.hard_reset();
Learn More
| Reference | Language | Description |
|---|---|---|
| Cactus Engine | C | Chat completion, streaming, tool calling, transcription, embeddings, RAG, vision, vector index, cloud handoff |
| Cactus Graph | C++ | Tensor operations, matrix multiplication, attention, normalization, activation functions |
| Cactus Kernels | C++ | ARM NEON SIMD kernels for matmul, attention, convolution, quantization, DSP, image processing |
| Cactus Quants | C++ | Rotation-and-codebook quantization from 4-bit to 1-bit for all weight tensors |
| Cactus Hybrid | C/Python | Route hard queries to the cloud automatically based on local model confidence |
| Cactus Transpiler | Python | Convert any PyTorch model to a Cactus runtime graph for on-device inference |
| Python Package | Python | Python package and CLI |
Build
cactus build --apple # iOS/macOS
cactus build --android # Android
cactus build --python # Python
cactus build # default static lib
Bindings
Model weights: Pre-converted weights for all supported models at huggingface.co/Cactus-Compute.
Benchmarks (CPU-only, no GPU)
- All weights INT4 quantised
- LFM: 1k-prefill / 100-decode, values are prefill tps / decode tps
- LFM-VL: 256px input, values are latency / decode tps
- Parakeet: 20s audio input, values are latency / decode tps
- Missing latency = no NPU support yet
| Device | LFM 1.2B | LFMVL 1.6B | Parakeet 1.1B | RAM |
|---|---|---|---|---|
| Mac M4 Pro | 582/100 | 0.2s/98 | 0.1s/900k+ | 76MB |
| iPad/Mac M3 | 350/60 | 0.3s/69 | 0.3s/800k+ | 70MB |
| iPhone 17 Pro | 327/48 | 0.3s/48 | 0.3s/300k+ | 108MB |
| iPhone 13 Mini | 148/34 | 0.3s/35 | 0.7s/90k+ | 1GB |
| Galaxy S25 Ultra | 255/37 | -/34 | -/250k+ | 1.5GB |
| Pixel 6a | 70/15 | -/15 | -/17k+ | 1GB |
| Galaxy A17 5G | 32/10 | -/11 | -/40k+ | 727MB |
| CMF Phone 2 Pro | - | - | - | - |
| Raspberry Pi 5 | 69/11 | 13.3s/11 | 4.5s/180k+ | 869MB |
Supported Transcription Model
- STT: 20s audio input on Macbook Air M3 chip
- Benchmark dataset: internal evals with production users
| Model | Params | End2End ms | Latency ms | Decode toks/sec | NPU | RTF | WER |
|---|---|---|---|---|---|---|---|
| UsefulSensors/moonshine-base | 61M | 361.35 | 182 | 262 | yes | 0.0180 | 0.1395 |
| openai/whisper-tiny | 39M | 232.03 | 137.38 | 581 | yes | 0.0116 | 0.1860 |
| openai/whisper-base | 74M | 329.37 | 178.65 | 358 | yes | 0.0164 | 0.1628 |
| openai/whisper-small | 244M | 856.79 | 332.63 | 108 | yes | 0.0428 | 0.0930 |
| openai/whisper-medium | 769M | 2085.87 | 923.33 | 49 | yes | 0.1041 | 0.0930 |
| openai/whisper-large-v3 | 1.55B | 5994 | 2050 | 15.72 | no | 0.2992 | - |
| nvidia/parakeet-ctc-0.6b | 600M | 201.77 | 201.44 | 5214285 | yes | 0.0101 | 0.0930 |
| nvidia/parakeet-tdt-0.6b-v3 | 600M | 718.91 | 718.82 | 3583333 | yes | 0.0359 | 0.0465 |
| nvidia/parakeet-ctc-1.1b | 1.1B | 279.03 | 278.92 | 4562500 | yes | 0.0139 | 0.1628 |
Supported LLMs
- Gemma weights are often gated on HuggingFace, needs tokens
- Run
huggingface-cli loginand input your huggingface token
| Model | Features |
|---|---|
| google/gemma-3-270m-it | completion |
| google/functiongemma-270m-it | tools |
| google/gemma-3-1b-it | completion, gated |
| google/gemma-4-E2B-it | vision, audio, completion, tools, Apple NPU |
| google/gemma-4-E4B-it | vision, audio, completion, tools, Apple NPU |
| google/gemma-3n-E2B-it | completion, tools |
| google/gemma-3n-E4B-it | completion, tools |
| Qwen/Qwen3-0.6B | completion, tools, embed |
| Qwen/Qwen3-Embedding-0.6B | embed |
| Qwen/Qwen3.5-0.8B | vision, completion, tools, embed |
| Qwen/Qwen3-1.7B | completion, tools, embed |
| Qwen/Qwen3.5-2B | vision, completion, tools, embed |
| LiquidAI/LFM2.5-350M | completion, tools, embed |
| LiquidAI/LFM2-700M | completion, tools, embed |
| LiquidAI/LFM2-8B-A1B | completion, tools, embed |
| LiquidAI/LFM2.5-1.2B-Thinking | completion, tools, embed |
| LiquidAI/LFM2.5-1.2B-Instruct | completion, tools, embed |
| LiquidAI/LFM2-2.6B | completion, tools, embed |
| LiquidAI/LFM2-VL-450M | vision, txt & img embed, Apple NPU |
| LiquidAI/LFM2.5-VL-450M | vision, txt & img embed, Apple NPU |
| LiquidAI/LFM2.5-VL-1.6B | vision, txt & img embed, Apple NPU |
| tencent/Youtu-LLM-2B | completion, tools, embed |
| nomic-ai/nomic-embed-text-v2-moe | embed |
Using this repo
┌──────────────────────────────────────────────────────────────────────────────┐
│ │
│ Step 0: if on Linux (Ubuntu/Debian) │
│ sudo apt-get install python3 python3-venv python3-pip cmake │
│ build-essential libcurl4-openssl-dev │
│ │
│ Step 1: clone and setup │
│ git clone https://github.com/cactus-compute/cactus && cd cactus │
│ source ./setup │
│ │
│ Step 2: use the commands │
│──────────────────────────────────────────────────────────────────────────────│
│ │
│ cactus auth manage Cloud API key │
│ --status show key status │
│ --clear remove saved key │
│ │
│ cactus run [model] chat playground (gemma-4-E2B-it) │
│ --image <path> image file for VLM inference │
│ --audio <path> audio file (WAV) for audio chat │
│ --system <prompt> system prompt │
│ --prompt <text> send prompt immediately │
│ --thinking enable thinking/reasoning mode │
│ --token <token> HF token (gated models) │
│ --reconvert force reconversion from source │
│ │
│ cactus transcribe [model] speech-to-text (parakeet-tdt-0.6b-v3) │
│ --file <audio.wav> audio file to transcribe (required) │
│ --language <code> language code (default: en) │
│ --token <token> HF token (gated models) │
│ --reconvert force reconversion from source │
│ │
│ cactus download <model> fetch pre-converted CQ from Cactus-Compute │
│ --bits 1|2|3|4 CQ quantization (default: 4) │
│ --token <token> HuggingFace API token │
│ │
│ cactus convert <model> [dir] convert model to CQ format │
│ (pre-converted if available, else │
│ built from source) │
│ --bits 1|2|3|4 CQ quantization (default: 4) │
│ --token <token> HuggingFace API token │
│ --reconvert force build from source │
│ │
│ cactus build build for ARM → build/libcactus.a │
│ --apple Apple (iOS/macOS) │
│ --android Android │
│ --python shared lib for Python FFI │
│ │
│ cactus test run unit tests and benchmarks │
│ --model <model> default: google/gemma-4-E2B-it │
│ --suite <name> run a specific suite (llm, vlm, stt, │
│ embed, rag, graph, index, kernel, etc.) │
│ --token <token> HuggingFace API token │
│ --reconvert force reconversion from source │
│ --ios run on connected iPhone │
│ --android run on connected Android │
│ │
│ cactus clean remove all build artifacts │
│ cactus --help show all commands and flags │
│ │
└──────────────────────────────────────────────────────────────────────────────┘
Maintaining Organisations
- Cactus Compute, Inc. (YC S25)
- UCLA's BruinAI
- Char (YC S25)
- Yale's AI Society
- National University of Singapore's AI Society
- UC Irvine's AI@UCI
- Imperial College's AI Society
- University of Pennsylvania's AI@Penn
- University of Michigan Ann-Arbor MSAIL
- University of Colorado Boulder's AI Club
Citation
If you use Cactus in your research, please cite it as follows:
@software{cactus,
title = {Cactus: AI Inference Engine for Phones & Wearables},
author = {Ndubuaku, Henry and Cactus Team},
url = {https://github.com/cactus-compute/cactus},
year = {2025}
}
N/B: Scroll all the way up and click the shields link for resources!