baya1116/deep-charger

Deep Charger — shipped on-device stack (iOS)

This repo is the complete stack of the "Deep Charger" iOS app: the shipped model weights AND the app's Swift project. Everything runs fully on-device (no server): MLX 4-bit LLM + SP hypernetwork pooler (bounded-memory long-context compression) + small task heads.

Contents

path	what it is
model/	LLM: DeepSeek-R1-Distill-Qwen-1.5B, full-fine-tuned (Phase15 FFT), MLX 4-bit
assets/pooler.safetensors	SP hypernet pooler (fp32 canonical) — compresses past tokens into 32 soft-prompt slots
assets/pooler_q4.safetensors	4-bit pooler actually loaded on iPhone
assets/pooler_config.json	pooler config
assets/intent_head.json, assets/specificity_head.json, assets/trigger_head.json	routing / WM-pinning / recall-trigger heads
assets/phaseb_indexer.safetensors, assets/recall_gate_v2.safetensors, assets/trigger_gate_v3.safetensors	retrieval / recall gates
app/	the iOS app's Swift project (mlx-swift port: Pooler.swift, SPQwen2.swift input-embeddings + cache-trim, SPGenerator.swift, tiered memory + recall triggers). Team ID redacted; build with your own signing.

Why the app source is here

The runtime/operational side (decode policy, SP eviction scheduling, recall triggering, translation UX) is as important as the weights — this repo keeps weights + operations in ONE place.

Note: mlx-swift does not run on the iOS Simulator (Metal); validate on macOS or a real iPhone GPU.

Architecture (1 line)

1.5B reasoner + recurrent attention-pooling hypernet: past context is evicted into 32 SP vectors (peak memory ~ f(batch), independent of context length) with a raw-token recency window; the app adds tiered memory + verbatim recall triggers on top.

Related repos

• baya1116/hypernet-sp-distill — research workspace (training runs, experiments; messy by design)
• baya1116/Phase15-DeepSeek-FFT — the FFT training phase this model came from