Roadmap

llamdrop develops in focused phases. Each version ships working features, not promises.

---

v0.3 — Done ✅

• One-command install, no compilation
• Prebuilt llama.cpp binary for Android
• Auto device detection — RAM, CPU, OS, storage
• Verified model catalog with tier system
• Live HuggingFace search
• Resilient downloader with resume and retry
• Smart quantization selection
• Session save and resume
• Background catalog update checker

---

v0.4 to v0.7 — Done ✅

• Phone-wide GGUF scanner
• Vulkan GPU acceleration
• Live RAM monitor and context trimming
• llamdrop update and llamdrop doctor
• Model benchmarking
• Config file
• Battery monitoring
• Chip-aware thread count
• Ollama backend auto-detect (Linux)
• Conditional mmap

---

v0.8.x — Done ✅

• Full DeviceProfile dataclass rollout
• Device-aware browser
• Cancelled downloads cleanup
• Incremental prompt buffer
• Arabic UI language added

---

v0.9.x — Done ✅

• Multi-language UI Support (English, Hindi, Spanish, Portuguese, Arabic)
• Optional File Context Injection (chat with PDFs, TXTs, CSVs)
• Unified RAM reading
• Fixed Inference UI Blocking
• Formal Testing Suite

---

v0.10.0 — Done ✅ (LTS Core Pivot)

• Dynamic Backend Probing: Auto micro-benchmark for best backend selection.
• Independent Engine Updates: Update llama.cpp separately.
• Auto-healing Doctor: [F] key auto-fixes broken configs/binaries.
• Power User Overrides: allow_thermal_melt, backend forcing, [U] unsupported models toggle.
• Dynamic Catalog Overhaul: Background catalog fetching and direct HF URL pasting.
• Dropped Windows/macOS: Focus exclusively on Android (Termux), Linux, and SBCs.
• Dropped Desktop Tier: Focused on accessible hardware from Micro to High (up to 16GB).

---

v1.0 — Planned

• Web-based model catalog on GitHub Pages — browse models before installing
• Community device profile submissions
• llamdrop server mode — run model on phone, access from browser on same WiFi
• Streaming tokens via Ollama backend