KevinTex 1.2.10

Windows startup reliability release.

• Fixes the Python.Runtime.Loader.Initialize crash seen after extracting a downloaded Windows archive.
• Ships the required .NET application configuration next to KevinTex.exe.
• Pins the Windows PyWebView/Python.NET packaging stack to compatible versions.
• Extends the packaged smoke test to import the real WinForms backend so this failure is caught during release builds.

The CPU Windows ZIP is recommended for broad compatibility. CUDA, Vulkan, Linux, and Apple Silicon macOS packages are also attached below.

KevinTex 1.2.9

KevinTex 1.2.9 is the complete cloud Voice-to-LaTeX and cross-desktop snip release.

Highlights:

• Dictate formulas with the cloud provider using Gemini 3.5 Flash and the same saved Google AI Studio API key used for cloud image recognition.
• Keep image recognition on Gemma 4 while routing voice to the audio-capable Gemini model.
• Disclose cloud audio handling and enforce a safe 14 MiB inline-audio limit.
• Fall back to gnome-screenshot, grim, Spectacle, or ffmpeg on Linux when Pillow cannot access the desktop.
• Keep the snip selector above the application window so it reliably receives drag input.
• Correct the Linux fallback regression test so it is deterministic on macOS release runners.

Local model weights are not bundled. Existing model downloads and provider settings are preserved across the upgrade.

KevinTex 1.2.8 (superseded)

This release was superseded during platform validation by KevinTex v1.2.9, which includes a deterministic macOS CI correction. Use v1.2.9.

KevinTex 1.2.7

KevinTex 1.2.7 adds cloud Voice-to-LaTeX through Google AI Studio and bundles the latest drafted improvements.

Highlights:

• Dictate formulas while using the cloud provider; WAV audio is sent to Gemini 3.5 Flash and returned as editable Markdown and LaTeX.
• Reuse the same saved Google AI Studio API key for image and voice recognition.
• Keep cloud image recognition on Gemma 4 while routing voice to the audio-capable Gemini model.
• Clearly disclose that cloud voice recordings are sent to Google.
• Enforce a 14 MiB cloud-audio limit before encoding to stay within the inline request limit.
• Include regression tests for cloud audio payloads, provider behavior, and upload safety.

Local Voice-to-LaTeX remains available through MLX on Apple Silicon and llama.cpp on Windows/Linux. Local model weights are not bundled, and existing downloads are preserved across the upgrade.

KevinTex 1.2.6

KevinTex 1.2.6 adds Google AI Studio as an alternative to downloading local
Gemma 4 weights, broadens Windows/Linux hardware support, and fixes the macOS
desktop experience.

Highlights:

• Choose local Gemma 4 weights or a Google AI Studio API key on first launch
  and later from Settings.
• Fix Retina Snip & Get scaling while preserving full-resolution crops.
• Retry and resume macOS MLX model downloads when Hugging Face returns an Xet
  bridge 401, using the stable HTTPS downloader by default.
• Add optimized CPU, NVIDIA CUDA, AMD ROCm/HIP, Vulkan, and Intel SYCL runtime
  selection. Windows release artifacts include CPU, CUDA, and Vulkan flavors;
  HIP and SYCL builds are supported locally with their vendor SDKs.
• Stop model/server resources when the native app closes and make provider
  switching safe during active inference.
• Store Linux models and settings in the current user's writable app-data
  directory and migrate existing downloads.

The local Gemma model weights are not bundled. Existing downloaded weights are
preserved across the upgrade.

KevinTex 1.2.5

Gemma 4 through Google AI Studio

KevinTex 1.2.5 adds a first-launch choice between downloading local Gemma 4 weights and using hosted Gemma 4 with a Google AI Studio API key. The same provider controls are available later in Settings.

Highlights

• Google AI Studio cloud recognition with gemma-4-26b-a4b-it
• API keys remain in a user-only local configuration file
• full-width API-key input and clearer provider status
• safe local/cloud switching with backend resource cleanup
• Linux app shutdown now stops the model and server
• writable per-user model storage for .deb installs, with automatic migration of existing weights
• updated Linux, Windows, and Apple Silicon macOS packages

Local Gemma and MLX remain available for fully on-device recognition and voice input. Images are sent to Google only when the AI Studio provider is selected.

KevinTex 1.2.4

What's new

• Voice-to-LaTeX on Windows, Linux, and Apple Silicon macOS
• Native Gemma 4 audio understanding through llama.cpp or MLX
• New Linux GitHub Actions release build
• MLX verification and model-loading fixes
• Startup error reporting and Pillow compatibility fixes

All recognition remains local and offline after model download.

KevinTex 1.2.3

Summary

• MLX load fix: OptiQ Gemma weights are no longer rewritten through NumPy, which caused data type 'bfloat16' not understood
• macOS smoke test: verifies mlx-vlm imports and that OptiQ caches are left untouched during startup
• Includes the v1.2.2 MLX model switch + dark mode changes

macOS install

Download KevinTex-1.2.3-macOS-arm64.dmg or .zip (Apple Silicon only).

If upgrading from 1.2.2, delete the old MLX cache if load still fails:

rm -rf ~/Library/Application\ Support/KevinTex/models/gemma-4-e2b-it-mlx-4bit
rm -rf ~/Library/Application\ Support/KevinTex/models/gemma-4-e2b-it-mlx-optiq-4bit

Then relaunch to re-download the OptiQ MLX model (~5.4 GB).

KevinTex 1.2.2

Summary

• macOS MLX fix: switches to mlx-community/gemma-4-e2b-it-OptiQ-4bit with corrected audio-tower weights (fixes the audio_tower.subsample_conv_projection shape error)
• Dark mode: header toggle plus Settings → Appearance (Light / Dark / Match system)
• Purges the broken legacy MLX cache on upgrade

macOS install

Download KevinTex-1.2.2-macOS-arm64.dmg or .zip from this release (Apple Silicon only).

On first launch after upgrading, KevinTex will re-download the MLX model (~5.4 GB).

KevinTex 1.2.1

Summary

• Restores the stable Gemma 4 E2B-it Q4_K_M model (~4.1 GB) as the default backend
• Fixes instruction-following so OCR returns transcription only, not analysis/commentary blocks
• Requires LATEX:-formatted output in fast mode and strips meta junk in post-processing
• Purges the superseded mobile QAT model cache on upgrade

Install

Ubuntu (.deb): download kevintex_1.2.1_all.deb below, then:

sudo apt install ./kevintex_1.2.1_all.deb

From source (current user):

packaging/install-user.sh

Restart KevinTex after upgrading so the server reloads the new code.

Test plan

• Unit tests for meta-junk stripping (19 passing)
• Recognize a textbook snippet — output should be transcription only
• Upgrade from 1.2.0 — old QAT cache removed, Q4_K_M re-downloaded if needed