smdesai/GLiDE-0.4B-CoreAI

GLiDE-0.4B CoreAI

CoreAI conversion of fastino/GLiDE-0.4B, pinned at 6bda7ab91aeb36dcad6d27828a8871e548c7fc71. This is a decision model returning probabilities for choice, yes/no (noul), and score questions. Built with Fastino GLiDE. It does not generate text. Repository: smdesai/GLiDE-0.4B-CoreAI.

Files

Path	Purpose
h19p/graphs/glide.aimodelc/	iOS h19p GPU compiled short graph, static buckets 128–4096
h19p/graphs/glide_long.aimodelc/	iOS h19p GPU compiled 7999-token graph
h19p/graphs/static_buckets.json	Function mapping required by the short graph
assets/tokenizer.json	Encoder and render tokenizer data for the native Swift host
assets/prompt.json	Publisher prompt wrapper and decision configuration
aimodel/*.aimodel/	Portable source graphs with debug locations stripped
aimodel/static_buckets.json	Portable graph function mapping
manifest.json, h19p/manifest.json	File sizes and SHA-256 digests; runtime download inventory
provenance.json	Upstream revision, precision policy, stripping and compilation checks
LICENSE, NOTICE, licenses/	Publisher license, notices and base model license

Both graphs and the bucket map must remain together with these filenames. The app downloads h19p/graphs and assets on first launch into its application support directory. Downloads resume after interruption, verify checksums, and place each completed item atomically. Subsequent launches use the cached files offline. Inference stays on device. The source checkpoint, Python environment, oracle fixtures, and conversion logs are unnecessary for the app. The runtime download is 5,034,723,779 bytes (4.69 GiB); the complete upload tree is about 6.62 GB.

Requirements and precision

Requires CoreAI on iOS 27 and an h19p device for these compiled assets, with the increased-memory-limit entitlement. Other architectures require their own compilation of the portable sources. The short graph uses FP16 embeddings and encoder projections with FP32 normalization, RoPE, attention, GELU, residuals and readout; the long graph uses 256-query chunks. The context limit is 7999 tokens including special tokens. Use the port's native GlideKit tokenizer, prompt renderer and decision API to preserve publisher semantics.

Validation

The original mixed-precision exports passed 73 reference decisions on macOS and iPhone Duo. The eight demo answers matched the pinned FP32 publisher with maximum probability difference 0.000194550. These are regression checks, not a population accuracy estimate. The shipping preparation checks equal operation counts before stripping, after stripping and on reload; equal CoreAI inspect summaries; hashes; compiled source identity; and all-file binary scans for personal paths. The stripped shipping assets have not undergone a new physical-device parity run. See the manifests for exact byte counts and provenance for the preparation record.

License and modifications

Use is governed by the publisher's LICENSE, retrieved from the upstream Hugging Face repository. The Ettin base model license is retained under licenses/. Modifications: CoreAI conversion, mixed precision, static short functions, chunked long attention, debug-location removal and iOS h19p GPU compilation. Upstream notices are retained in NOTICE.