Version Archive & Changelog

Changelog history and immutable releases

v1.5.0 - UltraFace SSD Deep Learning Detector & ONNX Runtime

Production Release (Current) 2026-10-01
  • UltraFace SSD Deep Learning Neural Face Detector: Replaced legacy 2001-era Haar Cascade with production UltraFace RFB Single-Shot Detector (SSD) via ONNX Runtime 1.27.0 ARM64 NEON.
  • 100% False-Positive Elimination (Zero Ghost Box): Completely eradicated ceiling light and white wall false detections (10 ghost boxes -> 0), achieving 100.0% confidence on human faces.
  • 11.94ms Real-Device Ultra-Low Latency: Validated 11.94ms inference latency (>80 FPS real-time tracking) on Samsung Galaxy S25 (Snapdragon 8 Elite).
  • 1927 Solvay Conference 29-Scientist Crowd Benchmark: Successfully localized 29 / 29 historic scientists with 98.8% accuracy (1623.4ms on Galaxy S25, 4890.1ms on Galaxy A53).
  • Automated 4-Axis Idempotent Asset Provisioning: 'termux-vision install' now automatically verifies python-onnxruntime and provisions verified UltraFace ONNX weights alongside NEON C++, Vulkan GPU, and VLM engines.
  • Enhanced CLI & API: Added '--backend' (auto/neural/haar), '--score-threshold', and confidence scoring to 'termux-vision detect-face'.

v1.4.6 - Native Vulkan Compute Canny & 5-Backend Standard

Production Release 2026-09-29
  • 100% Native Vulkan GPU Compute Canny Edge Pipeline: 3-pass SPIR-V shader chain within VRAM via vkCmdPipelineBarrier, achieving 0.23ms (834x speedup vs Python) on Snapdragon 8 Elite (Adreno 830).
  • Ultra-Fast ARM64 NEON C++ Kernel: Eliminated trigonometric atan2f via tangent ratio bit quantization and 1-byte direction buffers, running Canny filtering in 3.02ms on Snapdragon 865 and 4.36ms on Exynos 1380.
  • Prebuilt-Asset-First Idempotent Installer: New 'termux-vision install' command instantly provisions verified precompiled ARM64 native binaries (1.79MB archive) with SHA-256 validation; zero-build 0.005s instant skip if already installed.
  • Unified 5-Backend Architecture: Fully standardized ['auto', 'gpu', 'vulkan', 'opencl', 'cpu'] and shorthand flags (--gpu, --cpu, --opencl) across all subcommands.
  • Zero-Deception Fail-Fast Multimodal Gatekeeper: Enforced strict binary pre-validation rejecting defective text-only binaries lacking --mmproj (E015) and corrupted model weights (E014), permanently banning silent fallbacks.
  • Galaxy Fleet Multi-Device Validation: Confirmed 100% operational integrity on Galaxy S25 (Adreno 830), Galaxy S20 (Snapdragon 865), and Galaxy A35 (Exynos 1380).

v1.4.0 - Snapdragon 8 Elite Full-GPU VLM Acceleration

Production Release 2026-09-07
  • Qualcomm Snapdragon 8 Elite (Adreno 830) Full-GPU VLM acceleration verified on Galaxy S25 (15.00 tok/s generation on Moondream2 1.8B f16, 2,706 MiB VRAM).
  • SPIR-V compiler register allocation patch: mul_mat_vec_max_cols reduced from 8 to 2, resolving Qualcomm VK_ERROR_UNKNOWN JIT compiler aborts.
  • Android kernel GPU watchdog (kgsl) timeout (ErrorDeviceLost) defense: GGML_VULKAN_SKIP_CHECKS='999999999' injection and micro-batch prefill chunking (-b 64 -ub 64).
  • ZeroFlickerEngine parameter un-clamping: full thread and inference argument passthrough to VisionAdapter.
  • Universal installer (install.sh) audit: single-command wheel resolution preventing legacy PyPI package overwrites.

v1.3.0 - ARM Mali GPU Acceleration & CLI Standardization

Production Release 2026-09-07
  • Full standardization of CLI parameters (-d/--device, -b/--backend, -i, -p, -m, --mmproj, -n, -c, -t, -W, -H, --tune-mali, --json).
  • Verified ARM Mali GPU (G78 MP14, G68 MP5) acceleration pipeline with 0.00 MiB CPU Mapped VRAM.
  • Real-device benchmarks: Galaxy S21 12.65 tok/s (+58.9%), Galaxy A35 5.47 tok/s (+55.8%).
  • Dynamic GitHub Releases asset resolver and automated wheel installer with SHA-256 integrity verification.
  • Architecture roadmap defined: Adreno and Xclipse backends marked as In Development.

v1.2.0 - Native C/C++ Dual Acceleration Engines

Production Release 2026-09-01
  • Official ameva-vulkan-runtime dynamic bridge integration for truthful hardware doctor probing.
  • Native C/C++ dual acceleration engines with 8-directional BFS Hysteresis queue.
  • Full COCO-80 standard class alignment for TinyYOLONano detector.

v1.1.0 - Initial Dual Engine SDK Launch

Production Release 2026-08-15
  • Dual Engine Python and Node.js SDK initial release.
  • Heuristic Haar Cascade face candidate detector.
  • Model Cache Manager and GGUF VLM loader.