Repository navigation
Releases
v1.8.1
Full Changelog: v1.8.0...v1.8.1
v1.8.0
v1.7.3
v1.7.2
v1.7.1
Firebrat 1.6.1 — library deletion and iPad validation
Visible library deletion and physical iPad validation.
- Native Apple library now has a visible trash button and a centered Delete/Cancel confirmation. Local deletion keeps original files and server copies, clears saved progress, and stops playback of the deleted book.
- Native offline PDF/OCR, illustrated reading, six Kokoro voices, novel casting and Qwen/Gemma conversion were tested on an iPad (9th generation), iPadOS 27.2. The published PWA reader, zoom and narration also passed in actual iPad Safari.
- Measured Qwen 0.6B generation: CPU 1.42–1.48 s, Metal 12.40–13.45 s on the short circuit fixture. CPU remains the default. Qwen 1.7B passed the source guard but took 98.15 s. Detailed timings, conservative coverage rejections and raw fixture reports are in the native README on swift-native. These are small-fixture checks, not full-book accuracy or battery benchmarks.
PWA: https://subhajit-roy-partho.github.io/firebrat/ (server conversion only, saved-book offline playback). Original website remains under Documentation.
Downloads: signed Android APK and an UNSIGNED IPA. The IPA requires signing/provisioning before installation; it is not an App Store or TestFlight build. Model weights download once before offline conversion. Native Drive sync, semantic figure cropping, equation-aware rewriting and resumable background conversion remain incomplete. Live Google account consent has not been verified on Apple devices.
Full Changelog: v1.6.0...v1.6.1
Firebrat 1.6.0 — Android, Apple and web
Android, native Apple devices and the server-conversion PWA.
Includes offline Apple PDF/OCR conversion, six Kokoro voices, novel casting, optional Qwen/Gemma compilation, local package import and synchronized reading. Core native flows and model conversion were verified on an iPad (9th generation); the PWA sample, image zoom and playback were verified in iPad Safari. Detailed benchmarks and remaining platform limits are in frontend/FirebratNative/README.md on swift-native.
PWA: https://subhajit-roy-partho.github.io/firebrat/
The IPA is UNSIGNED and needs signing/provisioning before installation. The llama.cpp ZIP is a pinned build dependency, not an installer. Native Drive sync, semantic figure cropping, equation-aware rewriting and resumable background conversion remain incomplete. Apple Google account consent remains unverified. A visible library Delete button is included in the 1.6.1 patch.
Firebrat Server 0.1.3 — desktop conversion
Guided desktop setup for converting PDFs on your computer and reading the narrated books on your phone.
- Apple silicon Mac: download the unsigned DMG and drag FirebratServer to Applications.
- Linux x86-64: use the
.runinstaller, tarball, or Docker launcher below. - Choose an illustrated book or novel, then convert through the browser setup screen.
- Run an Ollama language model locally, or configure your own OpenAI-compatible LLM API. Speech generation always stays local.
- Connect your phone privately using Tailscale Serve. The setup guide includes hardware/model suggestions and update instructions.
Initial setup downloads runtimes/model weights. Complete the first conversion online before expecting offline operation. Keep your computer awake while serving books.
Validation: 58 backend/launcher tests passed; Apple silicon and Ubuntu 22.04 packaged startup checks passed. Published DMG and Linux installer checksums were verified, and the downloaded DMG passed its own smoke test. On an M4/24 GB Mac, cached offline Chatterbox CPU generation produced 4.64 seconds of narration in 15.74 seconds after a 13.19-second model load. The sample WAV and measurements are attached. This is a short speech/startup test, not a full-book or NVIDIA performance benchmark.
The Mac build is unsigned/not notarized. Intel macOS lacks compatible wheels for this native pipeline; use a Linux backend/VM. No paid Apple enrollment was performed.
Docker Hub images
The guided Linux conversion server is also published to subhajitroy/firebrat on Docker Hub.
- CPU:
docker pull subhajitroy/firebrat:server-v0.1.3-desktop-cpu - NVIDIA GPU:
docker pull subhajitroy/firebrat:server-v0.1.3-desktop-gpu(NVIDIA driver and Container Toolkit required) - Moving tags:
desktop-cpuanddesktop-gpu.
Download Firebrat-Server-Docker-Linux.sh below and run it with cpu or gpu to open the setup screen. Model weights download on first use; books and settings persist on the host. LLM APIs are optional, and voice generation always stays local.
Setup, hardware/model choices, updates and private Tailscale access.
Both images passed a real backend startup check. NVIDIA execution/performance still requires a hardware test. Native macOS downloads support Apple silicon and are unsigned/not notarized; native Intel Mac builds are unavailable with the required PyTorch wheels.
Firebrat Server 0.1.2 — superseded preview
This preliminary build is superseded by Firebrat Server 0.1.3.
Use the newer release for the portable Linux installer, final setup guide, and CPU/NVIDIA Docker images. This preview is retained for development history.
Firebrat Server 0.1.0 — superseded preview
This preliminary build is superseded by Firebrat Server 0.1.3.
Use the newer release for the portable Linux installer, final setup guide, and CPU/NVIDIA Docker images. This preview is retained for development history.