Skip to content

LLAMdrop v0.10.0 — LTS Core Pivot

Latest
Latest

Choose a tag to compare

View all tags
@DeVenLucaz DeVenLucaz released this 19 Jun 10:59
· 1 commit to main since this release
v0.10.0
c665877

What's New

This is the biggest architectural shift since LLAMdrop was born. v0.10.0 marks the official "LTS Core Pivot" — we've stripped everything that didn't belong and doubled down on what LLAMdrop was always meant to be: the best local AI runner for Termux, Linux, and Raspberry Pi.

🔥 New Features

Dynamic Backend Probing
LLAMdrop no longer guesses which backend to use based on hardcoded rules. On first launch, it runs a live 10-second micro-benchmark using a dummy model — racing CPU vs Vulkan — and permanently sets whichever backend produces tokens faster on your specific device. Run llamdrop recalibrate anytime to re-run it.

Category & Provider Filters
The model browser now has two live filters. Press C to cycle by category (Chat / Coding / Reasoning / Multilingual / Ultra-Fast) and P to cycle by provider (Google / Meta / Alibaba / Microsoft / DeepSeek / Mistral). Both filters stack. Finding the right model now takes seconds.

Wildcard Search & Magic Paste
The HuggingFace search has been completely overhauled. Paste a direct repo path like bartowski/Qwen3-4B-GGUF and LLAMdrop fetches it instantly — no browsing needed. Regular search now sorts by trending score and filters to GGUF-only results.

Power User Overrides
Every safety feature can now be bypassed. Set allow_thermal_melt: true in config to ignore thermal/battery warnings. Press U in the browser to unhide models that exceed your RAM. Pass raw llama.cpp flags directly via config. We are not Apple.

Auto-Heal Doctor
llamdrop doctor now has an [F] shortcut that automatically repairs your environment — reinstalls missing binaries, recreates config directories, redownloads a missing catalog, and requests Termux storage permissions.

Independent Engine Updates
A new ⚡ Update Engine option in the main menu updates your llama.cpp binary independently from the app itself — so you always have the fastest inference engine even if LLAMdrop hasn't pushed an update in months.

Offline Resilience
LLAMdrop now caches the model catalog, search results, and pages locally. If HuggingFace is unreachable, the app continues working with cached data and displays an Offline Mode indicator instead of crashing.

🗂️ Expanded Model Catalog

The catalog has been rebuilt from scratch — 26 models across 5 tiers, every single one tagged with provider and category. Stale and superseded models have been pruned. New additions include Qwen3 0.6B, and the entire catalog now auto-refreshes in the background every 24 hours without requiring an app update.

🧹 What We Removed

  • Full Windows and macOS support (installers, code paths, CI workflows)
  • Desktop and Workstation RAM tiers (24GB+, 64GB+)
  • Hardcoded chip-name decision trees for backend selection
  • PowerShell installer (install.ps1)
  • Stale and superseded models (TinyLlama, Phi-3, Gemma 2 2B, and others)

🐛 Bug Fixes

  • Browser crash when opening model catalog (AttributeError: list object has no attribute items)
  • Main menu header showing Unknown - 0GB RAM instead of real device info
  • Dummy benchmark model appearing in "Start chatting" model list
  • Backend prober timeout crash on slow devices
  • IDX_ENGINE undefined crash when selecting Update Engine
  • prober.py missing from OTA update file list
  • Model catalog loading stale cached version instead of bundled update
  • best_for field displaying letter-by-letter in model detail panel
  • Test suite referencing removed Tier.DESKTOP and Tier.WORKSTATION

📦 Installation

curl -fsSL https://raw.githubusercontent.com/DeVenLucaz/llamdrop/main/install.sh | bash
Assets 2
Loading