Chiko Core ML models for iOS
Four optional model artifacts prepared for Chiko. These are model files, not the application source or user data. Initial physical testing is limited to iPhone 18 Pro (iPhone19,2). No general Japanese-quality or battery-saving guarantee is made.
| Prefix | Model | Default permitted compute | Limits |
|---|---|---|---|
| e5-512 | intfloat/multilingual-e5-small, local FP16 conversion | GPU | 512 tokens, 384 normalized dimensions, query/passage prefixes |
| whisper | Argmax Whisper Tiny Core ML | Neural Engine | Short synthetic Japanese fixture only; noise/name accuracy not established |
| qwen3vl | Qwen3-VL 2B Core ML | NE vision + GPU language | 1 image, 448 input, 2048 context, greedy; slow replies; matching Increased Memory Limit profile and >=5.5 GB process headroom required |
| sd-turbo | SD-Turbo Core ML, 256 preset | NE text + GPU denoiser/decoder | 256 square, 1-8 steps (default 2), CFG=0; prompt fidelity imperfect; no dedicated image-content classifier included |
Compute settings permit the device alongside CPU; they are not utilization measurements. Existing app safety and resource checks remain required. Entirely NE-based vision failed to load; entirely NE-based image generation had additional visual degradation. Qwen3 TTS and Core ML chat weights are not included.
Files use a flat download format. The prefix before -- selects the model. The remaining file name encodes package directories with __; Chiko reconstructs the original paths only after SHA-256 validation. Keep the artifact fileName and pathInRepo distinct. Do not load a GGUF into this format. File hashes and layout are in model-manifest.json.
Source and changes
- E5: intfloat/multilingual-e5-small revision 614241f622f53c4eeff9890bdc4f31cfecc418b3. Converted with coremltools 9.0, fixed 512 tokens, FP16; pooled embeddings are normalized by the app.
- Whisper: argmaxinc/whisperkit-coreml revision 0f63a7800b00dd0226abd051b906c246e1907482, openai_whisper-tiny; offline tokenizer files included.
- Vision: mlboydaisuke/qwen3-vl-2b-stateful-coreml revision fd430e92fc862fba5f4427513927e0bca48fece8, based on Qwen/Qwen3-VL-2B-Instruct; original model math retained, loading is split across NE-permitted image encoder and GPU-permitted language. See per-file hashes for this selected layout.
- SD-Turbo: stabilityai/sd-turbo revision b261bac6fd2cf515557d5d0707481eafa0485ec2, converted with Apple ml-stable-diffusion tooling to 256 square with split UNet. GPU rendering is used to avoid observed extra all-NE degradation. Compiled files were tested on the stated device only.
License and attribution
This repository combines separately licensed models; it is not all Apache-2.0 or all MIT. See LICENSES.md and the included original license texts. Preserve them and the applicable notices on redistribution.
Powered by Stability AI. SD-Turbo is subject to Stability AI Community License Agreement and its Acceptable Use Policy. Commercial use may require registration or an enterprise agreement under the original terms. The distributor's approval does not accept these terms for downstream users. Chiko must collect the user's required consent before download.
Generated text/images may be wrong or unsuitable. Do not rely on outputs for consequential decisions. This collection is provided under its original licenses without additional quality promises.
- Downloads last month
- -