pinkelephantlimited

phone-agent

Why Own Model

We learned from the whole HF universe (not just pinkelephantlimited):

PocketPal AI / Maid / MLC Chat (2026) — GGUF Q4_K_M is the community standard; base APK ~20MB, models downloadable, Vulkan/CPU universal. We adopt llama.cpp via NDK + GGUF (already proven in PinkElephantAgent) for max portability, not vendor-locked LiteRT/ML Kit.
HF RAG Spaces — best pattern is local-first, not local-only: sequential local RAG + web search, concatenated with section headers ([Local], [Web Search Results]) so LLM knows source. Dark mode retrieval discussions confirm structure-aware chunking + 384d embedders beat naive splitting.
PromptQuorum / ItsFOSS 2026 tests — Snapdragon 8 Elite ~40 tok/s (NPU), Snapdragon 870 (iQOO Neo5 SE) ~12-18 tok/s CPU/Vulkan on Phi-3-mini/Gemma2B class. Our 1.2B Q4_K_M targets ~15 tok/s on your device — DeepSeek-like streaming, no cloud.

We do not bundle Qwen. Own tokenizer + own 1.2B config (phone_llm_train/model_config.json) ensures brand sovereignty.

Architecture: Local + Web, DeepSeek UI

markdown
User → WebSearchService (Jina Reader + DDG lite, 8s timeout, capped 1500 chars)
     └→ local prompt (PhoneLLM GGUF via NativeEngine.nativeLlamaChat, 2048 ctx, 4 threads)
        → DeepSeek-like RecyclerView (streaming bubble, markdown, history, web toggle)

Base APK ~20MB — models not bundled (learned from Pocket LLM v1.4: 200MB base → downloadable). APK lives in apk/ on HF for storage.
Storage: app-private getExternalFilesDir()/models — correct scoped storage (PocketPal pattern).
Training vehicle: molab RTX PRO 6000 102GB (sb-bf9f066c447abaab), BF16, PagedAdamW8bit, gradient checkpointing — same QLoRA lineage as 48B but from scratch.

The Phone Model at a Glance

Table with columns: Spec, Value
Spec	Value
Params	1.2B dense, 24L, 16 heads, 4 KV, hidden 2048, intermediate 5504, vocab 32000, RoPE 10k, 4096 ctx
Tokenizer	Own BPE 32k (`[PAD][BOS][EOS][WEB][THINK]`), trained on FineWeb-Edu + OpenHermes + StarCoder (HF datasets)
Quant	Q4_K_M ~0.8GB (phone), Q4_0 ~0.6GB (low-end), F16 ~2.4GB
Teacher	Pink Elephant 48B-S (47.7B) — distillation, not copy
Dataset	HF permissive mix 60% web / 30% chat / 10% code — curation is the moat
Hardware	molab 102GB → iQOO Neo5 SE 8GB offline
Inference

Repo Layout (HF as storage)

markdown
pinkelephantlimited/phone-llm/
  README.md
  config.json                 # own 1.2B config (sovereign)
  tokenizer.json              # own BPE (after train)
  train_tokenizer.py          # from-scratch BPE
  train_pretrain.py           # 1.2B pretrain on molab
  OWN_MODEL_PLAN.md
  phone-llm-Q4_K_M.gguf       # quantized for APK (after training)
  phone-llm-f16.gguf
  training/phone-llm/         # logs, loss curves
  apk/phone-llm-v1.0.apk      # base APK

Quick Start (on-device)

Install APK from apk/phone-llm-v1.0.apk via adb install
First launch → taps Download model (HF) → fetches phone-llm-Q4_K_M.gguf to models/ with progress
Toggle Web Search for hybrid mode

bash
adb install apk/phone-llm-v1.0.apk
# or via HF
huggingface-cli download pinkelephantlimited/phone-llm phone-llm-Q4_K_M.gguf --local-dir ./models

Build From Source

bash
cd app/PhoneLLM
./gradlew assembleDebug  # ~20MB base APK
adb install app/build/outputs/apk/debug/app-debug.apk

Training (molab GPU)

bash
python phone_llm_train/train_tokenizer.py
python phone_llm_train/train_pretrain.py  # 102GB GPU, 5000 steps v0.1
python -m llama_cpp.convert_hf_to_gguf --outfile phone-llm-Q4_K_M.gguf --outtype q4_k_m /home/marimo/phone-llm-out/final

APK

See app/PhoneLLM — DeepSeek-like chat, Hybrid RAG, HF storage. Test on iQOO Neo5 SE via adb/scrcpy.

Pink Elephant Limited — sovereign on-device intelligence, 2026

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

pinkelephantlimited

Model Tree

Base

pinkelephantlimited/phone-llm

Fine-tuned

this model

Input Modalities

Text

Output Modalities