Remove coreml

README.md

@@ -124,7 +124,7 @@ graph.hard_reset();
 
 - LLM: Gemma-4-E2B-CQ4 (1k-context prefill / decode for 100 tokens)
 - VLM: Gemma-4-E2B-CQ4 (256px image encode time / decode)
-- Transcribe: Parakeet-TDT-0.6B-CQ4 (20s audio ens-to-end transcribe time)
+- Transcribe: Parakeet-TDT-0.6B-CQ4 (20s audio end-to-end transcribe time)
 - 1k-Context RAM: peak MB during the LLM benchmark
 
 Command: `cactus benchmark` [optional `--ios` or `--android`]
@@ -208,8 +208,7 @@ Command: `cactus benchmark` [optional `--ios` or `--android`]
 │                                                                                │
 │  cactus run [model|path]             run a model (downloads if needed)         │
 │    --bits 1|2|3|4                    CQ quantization (default: 4)              │
-│    --weights general|apple           weights bundle variant (default: general) │
-│    --backend auto|cpu|metal          inference backend (default: auto)         │
+│    --backend cpu|metal               inference backend (default: auto)         │
 │    --image <path>                    image file for VLM inference              │
 │    --audio <path>                    audio file for audio chat                 │
 │    --system <prompt>                 system prompt                             │
@@ -222,20 +221,17 @@ Command: `cactus benchmark` [optional `--ios` or `--android`]
 │    --file <audio.wav>                audio file to transcribe (WAV)            │
 │    --language <code>                 language code (default: en)               │
 │    --bits 1|2|3|4                    CQ quantization (default: 4)              │
-│    --weights general|apple           weights bundle variant (default: general) │
 │    --token <token>                   HuggingFace token (gated models)          │
 │    --reconvert                       force local rebuild from source           │
 │                                                                                │
 │  cactus download [model]             get a bundle (prebuilt, else build)       │
 │    --bits 1|2|3|4                    CQ quantization (default: 4)              │
-│    --weights general|apple           weights bundle variant (default: general) │
 │    --token <token>                   HuggingFace token (gated models)          │
 │    --reconvert                       force local rebuild from source           │
 │                                                                                │
 │  cactus convert <model> [dir]        HuggingFace -> runnable cactus bundle     │
 │                                      (CQ weights + runtime graph)              │
 │    --bits 1|2|3|4                    CQ quantization (default: 4)              │
-│    --weights general|apple           weights bundle variant (default: general) │
 │    --token <token>                   HuggingFace token (gated models)          │
 │    --reconvert                       force local rebuild from source           │
 │    --lora <path>                     merge a LoRA adapter before converting    │
@@ -246,8 +242,7 @@ Command: `cactus benchmark` [optional `--ios` or `--android`]
 │    --host <addr>                     bind address (default: 127.0.0.1)         │
 │    --port <port>                     port (default: 8080)                      │
 │    --bits 1|2|3|4                    CQ quantization (default: 4)              │
-│    --weights general|apple           weights bundle variant (default: general) │
-│    --backend auto|cpu|metal          inference backend (default: auto)         │
+│    --backend cpu|metal               inference backend (default: auto)         │
 │    --token <token>                   HuggingFace token (gated models)          │
 │    --reconvert                       force local rebuild from source           │
 │    --no-cloud-handoff                disable automatic cloud handoff           │
@@ -257,8 +252,7 @@ Command: `cactus benchmark` [optional `--ios` or `--android`]
 │  cactus code                         run the AI coding agent (TUI / print)     │
 │    --serve-model <id>                auto-start a server with this model       │
 │    --bits 1|2|3|4                    CQ quantization (default: 4)              │
-│    --weights general|apple           weights bundle variant (default: general) │
-│    --backend auto|cpu|metal          inference backend (default: auto)         │
+│    --backend cpu|metal               inference backend (default: auto)         │
 │    --token <token>                   HuggingFace token (gated models)          │
 │    --reconvert                       force local rebuild from source           │
 │    --host <addr>                     server address (default: 127.0.0.1)       │
@@ -282,8 +276,7 @@ Command: `cactus benchmark` [optional `--ios` or `--android`]
 │    --model <hf-id>                   default: google/gemma-4-E2B-it            │
 │    --transcription-model <hf-id>     default: nvidia/parakeet-tdt-0.6b-v3      │
 │    --bits 1|2|3|4                    CQ quantization (default: 4)              │
-│    --weights general|apple           weights bundle variant (default: general) │
-│    --backend auto|cpu|metal          inference backend (default: auto)         │
+│    --backend cpu|metal               inference backend (default: auto)         │
 │    --token <token>                   HuggingFace token (gated models)          │
 │    --reconvert                       force local rebuild of test models        │
 │    --suite <name>                    run a single test suite by name           │
@@ -298,8 +291,7 @@ Command: `cactus benchmark` [optional `--ios` or `--android`]
 │    --model <hf-id>                   default: google/gemma-4-E2B-it            │
 │    --transcription-model <hf-id>     default: nvidia/parakeet-tdt-0.6b-v3      │
 │    --bits 1|2|3|4                    CQ quantization (default: 4)              │
-│    --weights general|apple           weights bundle variant (default: general) │
-│    --backend auto|cpu|metal          inference backend (default: auto)         │
+│    --backend cpu|metal               inference backend (default: auto)         │
 │    --ios                             run on connected iPhone                   │
 │    --android                         run on connected Android                  │
 │                                                                                │

apple/CMakeLists.txt

@@ -27,7 +27,7 @@ if(BUILD_SHARED_LIBS)
     )
 
     target_link_libraries(cactus_framework PUBLIC
-        "-framework Accelerate" "-framework CoreML" "-framework Foundation" "-framework Metal" "-framework MetalPerformanceShaders"
+        "-framework Accelerate" "-framework Foundation" "-framework Metal" "-framework MetalPerformanceShaders"
         "-framework Security" "-framework SystemConfiguration" "-framework CFNetwork"
     )
 

bindings/rust/cactus.rs

@@ -7,7 +7,8 @@ pub type CactusTokenCallback = Option<unsafe extern "C" fn(token: *const c_char,
 pub type CactusLogCallback = Option<unsafe extern "C" fn(level: c_int, component: *const c_char, message: *const c_char, user_data: *mut c_void)>;
 
 #[cfg_attr(target_os = "macos", link(name = "Accelerate", kind = "framework"))]
-#[cfg_attr(target_os = "macos", link(name = "CoreML", kind = "framework"))]
+#[cfg_attr(target_os = "macos", link(name = "Metal", kind = "framework"))]
+#[cfg_attr(target_os = "macos", link(name = "MetalPerformanceShaders", kind = "framework"))]
 #[cfg_attr(target_os = "macos", link(name = "Foundation", kind = "framework"))]
 #[cfg_attr(target_os = "macos", link(name = "Security", kind = "framework"))]
 #[cfg_attr(target_os = "macos", link(name = "SystemConfiguration", kind = "framework"))]

blog/gemma4.md

@@ -19,14 +19,14 @@ This matters because latency compounds. A pipeline that transcribes audio, then
 
 ## Performance
 
-Cactus targets ARM across platforms: Apple Silicon Macs, iPhones, iPads, Vision Pro, and Android devices with ARM64 chipsets (Snapdragon, Dimensity, Tensor). On Apple hardware we additionally leverage the Neural Engine for vision and audio encoders, while on Android and Linux the runtime uses NEON, i8mm, and dot-product intrinsics.
+Cactus targets ARM across platforms: Apple Silicon Macs, iPhones, iPads, Vision Pro, and Android devices with ARM64 chipsets (Snapdragon, Dimensity, Tensor). On Apple hardware we additionally leverage Metal GPU acceleration, while on Android and Linux the runtime uses NEON, i8mm, and dot-product intrinsics.
 
 | Metric (M5 Mac/iPad/Vision Pro) | E2B |
 |---|---|
 | 4096-token prefill | 660 tok/s |
 | 1024-token decode | 40 tok/s |
 | 30s audio end-to-end | 0.3s |
-| Image encode (ANE) | 0.7s |
+| Image encode | 0.7s |
 
 40 tok/s decode is faster than most people read. You're not waiting for the model. The model is waiting for you.
 
@@ -128,7 +128,7 @@ Cactus supports React Native, Flutter, Swift, Kotlin, Python, Rust, and C++. Pic
 |---|---|---|
 | React Native | [React Native bindings](/bindings/react-native/) | Native bridge over the C API |
 | Flutter | [cactus-flutter](/bindings/flutter/) | Dart FFI bindings |
-| Swift (iOS/macOS) | [cactus-swift](/bindings/swift/) | XCFramework with NPU support |
+| Swift (iOS/macOS) | [cactus-swift](/bindings/swift/) | XCFramework with Metal support |
 | Kotlin (Android) | [cactus-kotlin](/bindings/kotlin/) | JNI + Kotlin Multiplatform |
 | Python | [cactus-compute](/python/) | `pip install cactus-compute` |
 | Rust | [Rust bindings](/bindings/rust/) | Copy `cactus.rs`, link `libcactus_engine.a` |

blog/hybrid_transcription.md

@@ -53,7 +53,7 @@ cactus transcribe --file path/to/audio.wav
 ## Built for the Edge, from the Ground Up
 Achieving this required relying on our industry-leading on-device engine. We built Cactus from the ground up exactly for this.
 
-Because the engine sits directly on the metal, it is optimized for edge constraints. It targets ARM-based CPUs and NPUs natively, resulting in super low RAM usage (e.g. Moonshine transcription for example, runs with sub-10MB RAM utilization). This is a critical factor for mobile and wearable devices where memory is strictly rationed by the OS.
+Because the engine sits directly on the metal, it is optimized for edge constraints. It targets ARM-based CPUs natively, resulting in super low RAM usage (e.g. Moonshine transcription for example, runs with sub-10MB RAM utilization). This is a critical factor for mobile and wearable devices where memory is strictly rationed by the OS.
 
 By keeping the footprint minimal, the engine operates silently in the background without draining battery or causing OS-level memory warnings, making it viable for continuous, ambient listening.
 

blog/lfm2.5_350m.md

@@ -41,7 +41,7 @@ most people read at roughly 4-5 words per second, so even the Raspberry Pi's 30
 Second, the RAM gap between Apple and Android: 56-85 MB versus 300-330 MB for the same 355 MB model.
 That gap is not a bug. It's a direct consequence of how the two platforms handle memory-mapped files, which we'll get into in the Cactus format section below.
 
-## Why CPU, Not GPU or NPU
+## Why CPU, Not GPU or Dedicated Accelerators
 
 It is tempting to reach for the GPU. On paper, mobile GPUs have 10-50x the FLOPS of a single CPU core. 
 But inference on a 350M-parameter model is not a FLOPS problem, it's a memory bandwidth problem. 
@@ -55,17 +55,17 @@ That's two separate shader codepaths to maintain, with different memory models,
 And GPU inference holds a persistent GPU context, which competes with the UI rendering pipeline and drains battery even when the inference workload is light.
 GPUs are also energy-inefficient. On Macs and PCs this barely matters since they're plugged into a power source most of the time, and GPU inference works well there. But mobile devices run on battery, and every watt spent on GPU inference cuts directly into the user's runtime. The GPU path that works fine on a MacBook becomes a battery killer on a phone.
 
-NPU is a more interesting target. Apple's Neural Engine and Qualcomm's Hexagon DSP both offer high-throughput inference at good power efficiency. 
-The problem is access. Apple's Neural Engine requires CoreML model conversion, which imposes its own quantization and graph constraints.
+Dedicated AI accelerators are a more interesting target. Apple's Neural Engine and Qualcomm's Hexagon DSP both offer high-throughput inference at good power efficiency. 
+The problem is access. Apple's Neural Engine requires conversion into a proprietary model format, which imposes its own quantization and graph constraints.
 You don't control the execution schedule, the memory layout, or the precision semantics. 
-The quantisation techniques available through CoreML severely lag the state-of-the-art methods used by GGML and Cactus.
+The quantisation techniques available through that toolchain severely lag the state-of-the-art methods used by GGML and Cactus.
 Qualcomm's QNN SDK has similar restrictions (limited quantisation scheme support) and not all Android devices ship Qualcomm chips. 
-Both lock you into a vendor-specific format that may not support your model architecture at all, and neither works on budget devices that lack the NPU entirely.
+Both lock you into a vendor-specific format that may not support your model architecture at all, and neither works on budget devices that lack the accelerator entirely.
 
 CPU is the universal target. Every ARM device, flagship, budget, wearable, Raspberry Pi, and most recent ones support DOTPROD or I8MM extensions. 
-A single well-optimized CPU codepath works everywhere, the main challenge is energy-inefficiency compared to NPU.
+A single well-optimized CPU codepath works everywhere, the main challenge is energy-inefficiency compared to dedicated accelerators.
 
-ARM is also closing this gap. The progression from NEON to DOTPROD to I8MM to SME2 reflects a deliberate push to bring matrix-processing efficiency onto the CPU itself. Each generation narrows the energy and throughput gap with dedicated NPUs. SME2, landing on upcoming ARMv9.2-A cores, adds streaming matrix operations that approach NPU-class efficiency for INT8 workloads while remaining fully programmable. The bet is that general-purpose CPU silicon will continue absorbing the capabilities that once justified a separate accelerator.
+ARM is also closing this gap. The progression from NEON to DOTPROD to I8MM to SME2 reflects a deliberate push to bring matrix-processing efficiency onto the CPU itself. Each generation narrows the energy and throughput gap with dedicated accelerators. SME2, landing on upcoming ARMv9.2-A cores, adds streaming matrix operations that approach accelerator-class efficiency for INT8 workloads while remaining fully programmable. The bet is that general-purpose CPU silicon will continue absorbing the capabilities that once justified a separate accelerator.
 
 ## Why Single Core for Decode
 

cactus-engine/CMakeLists.txt

@@ -16,7 +16,6 @@ set(ENGINE_SOURCES
     src/constraints.cpp
     src/model.cpp
     src/kv_compress.cpp
-    src/model_npu.cpp
     src/engine_image.cpp
     src/index.cpp
     src/index_ffi.cpp
@@ -32,14 +31,6 @@ set(ENGINE_SOURCES
     src/stream.cpp
     src/tokenizer.cpp
 )
-if(APPLE)
-    enable_language(OBJCXX)
-    list(APPEND ENGINE_SOURCES src/npu_ane.mm)
-    set_source_files_properties(src/npu_ane.mm PROPERTIES COMPILE_FLAGS "-fobjc-arc")
-else()
-    list(APPEND ENGINE_SOURCES src/npu.cpp)
-endif()
-
 add_library(cactus_engine STATIC ${ENGINE_SOURCES})
 set_target_properties(cactus_engine PROPERTIES
     POSITION_INDEPENDENT_CODE ON
@@ -56,7 +47,7 @@ target_link_libraries(cactus_engine PUBLIC cactus_graph)
 
 if(APPLE)
     target_link_libraries(cactus_engine PUBLIC
-        "-framework CoreML" "-framework Foundation" "-framework Metal" "-framework MetalPerformanceShaders"
+        "-framework Foundation" "-framework Metal" "-framework MetalPerformanceShaders"
         "-framework Security" "-framework SystemConfiguration" "-framework CFNetwork"
     )
 
@@ -156,7 +147,7 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
             target_link_libraries(cactus_engine_shared PRIVATE "${_curl}")
         endif()
         target_link_libraries(cactus_engine_shared PRIVATE
-            "-framework Accelerate" "-framework CoreML" "-framework Foundation" "-framework Metal" "-framework MetalPerformanceShaders"
+            "-framework Accelerate" "-framework Foundation" "-framework Metal" "-framework MetalPerformanceShaders"
             "-framework Security" "-framework SystemConfiguration" "-framework CFNetwork"
         )
     else()

cactus-engine/cactus_engine.h

@@ -110,6 +110,17 @@ CACTUS_FFI_EXPORT int cactus_transcribe(
     size_t pcm_buffer_size
 );
 
+CACTUS_FFI_EXPORT int cactus_preprocess_audio_features(
+    const char* audio_file_path,
+    const char* model_type,
+    size_t mel_bins,
+    float* features_buffer,
+    size_t buffer_size,
+    size_t* feature_count,
+    size_t* out_mel_bins,
+    size_t* out_frames
+);
+
 CACTUS_FFI_EXPORT cactus_stream_transcribe_t cactus_stream_transcribe_start(
     cactus_model_t model,
     const char* options_json