Bambi Local alpha.41 — AI, voice & guided setup

Windows installer · Windows ZIP · Ubuntu 24.04 installer · Linux AppImage · Apple Silicon Mac

You only need one app download for your system. Choose Windows EXE or Windows ZIP; Ubuntu 24.04 uses the Debian installer; other compatible Linux systems use the AppImage; Apple Silicon Mac uses the Mac ZIP. The technical/source ZIP and checksums are optional for ordinary installation.

Install the Windows EXE, or extract the complete Windows ZIP and open Bambi Local.exe. On Mac (M series, macOS 14+), extract the ZIP and copy Bambi Local to Applications. On Ubuntu 24.04, install the Debian package with the system package installer, or sudo apt install ./bambi-local-0.1.0-alpha.41-linux-amd64.deb. It installs the Chromium sandbox correctly; the portable AppImage failed normal startup on the tested Ubuntu host. On Fedora, the AppImage passed the checks below. No Node, Python or compiler installation is needed.

AI and voice setup: Open the app, save Privacy & access, then choose Get AI & voice models in Setup. Review the model and choose Download and verify; the app downloads the AI, voice and English microphone data together. After verification, choose Start local AI. The reference 9B model and speech files need about 6.05 GB; allow 15 GB free disk space. Existing supported model files/folders can also be imported.

What improved since the previous releases

Comparison: Windows/Linux alpha.23 and the Mac-only alpha.29 tester.

Performance, memory and different PCs

• RAM-aware runtime profiles: compact, balanced and roomy systems receive different cache, checkpoint and batch allocations, with model-specific minimums and room for speech/the app. The selected companion instructions and full 24,576-token model context are retained.
• Better prompt-cache reuse: fixed companion instructions and heard history are kept separate from changing turn guidance, so a new turn does not needlessly invalidate the same reusable prompt prefix.
• Real performance calibration on your PC: Test performance compares eligible CPU/GPU paths, CPU thread counts and batch sizes. It records startup, first visible text, generation rate, memory and optional voice-synthesis timing. The initial full-prompt turn is separated from warm follow-up turns.
• Verified measured settings: a candidate configuration must pass a fresh-load repeat before being saved. Profiles stay local, match hardware/driver/runtime/model/prompt/voice settings and are rechecked when stale. Compare chosen models tests existing verified files and restores the original model selection.
• Smaller model choices: Hauhau Qwen3.5 4B (2.71 GB) and 2B (1.27 GB), plus CPU-focused LFM2.5 1.2B (0.73 GB), are available as experimental desktop options. Recommendations account for smaller-memory/shared-GPU systems. The reference 9B remains the default; smaller models can be less reliable with the full Classic personality.
• Acceleration and fallback improvements: Windows/Linux include CPU, Vulkan and both CUDA 12/13 runtime paths. Selection distinguishes device discovery from verified GPU layer offload; eligible startup/driver/memory failures can fall back to another path or CPU. Available-memory checks refine CPU admission.
• Explicit Linux Q8 KV-cache comparison: Hauhau 9B can try the experimental Q8 cache, with standard cache still the default. Change it only after stopping active work. DFlash/MTP remain Linux source/development experiments, not qualified desktop speedups.

These changes improve adaptation and expose actual measurements; there is no universal speedup claim for every PC. Full NVIDIA and Windows/Mac reference-model hardware checks remain open.

Setup, platforms and everyday use

• Easier AI and voice setup: Get AI & voice models opens the in-app flow for AI, Bambi Mix voice and English microphone files. Clearer progress, cancellation, retry and permission/error messages reduce manual model hunting.
• Windows installation fixes: storage upgrades, long installation paths/license notices and bounded hardware discovery were hardened; an optional portable ZIP is available alongside the EXE.
• Ubuntu 24.04 installation: a Debian installer sets up Chromium's sandbox correctly. Its installed application passed 22 checks; the AppImage remains the tested Fedora option.
• One current application across platforms: Apple Silicon Mac now receives the same model, memory, performance and profile features as the Windows/Linux release, with its native arm64/Metal and macOS permission handling.
• Profiles & backups: offline shareable setting codes and optional password-encrypted backups are added alongside Settings & presets. Export a backup before upgrading.
• Expanded settings transfer: the new model/cache and session controls are included in validated settings. Revision-checked saves, selected-section imports, named presets and undo remain available.
• Voice and captions: sentence delivery, cancellation and caption/playback handling were refined. Our original Bambi Mix through Kokoro remains the default voice.
• Runtime recovery and cleanup: model switching, cancellation and owned-process shutdown were hardened. Final Windows, Linux and Mac packages were checked for restart and cleanup behavior.
• More verification and easier access: 657 Node 22 tests, Node 24/browser/setup/settings checks, native Windows/Mac speech checks, Linux 9B GPU and small-model CPU conversations, and a spoken Linux restart without networking. All seven downloads were checked without sign-in and complete received files matched their SHA-256 checksums.

Models, memory and performance details.

Why are the downloads larger?

Approximate compressed download sizes, using decimal MB/GB:

App download	Previous platform release	Alpha 41
Windows installer	272 MB (alpha.23)	1.57 GB
Linux AppImage	356 MB (alpha.23)	2.03 GB
Apple Silicon Mac ZIP	322 MB (alpha.29)	322 MB

• The main increase is bundled NVIDIA GPU libraries. Windows and Linux now include both CUDA 12 and CUDA 13 runtime libraries for different NVIDIA driver generations, alongside CPU/Vulkan paths. Those libraries alone occupy about 1.83 GB on Windows and 2.15 GB on Linux before compression. The old Windows/Linux packages did not contain these CUDA folders. Actual NVIDIA hardware qualification is still open.
• These are runtime libraries, not the AI model weights. The selected AI, voice and microphone models remain a separate explicit in-app download. The reference 9B setup still downloads about 6.05 GB, as in alpha.23.
• Mac does not contain the CUDA libraries, so its download size is almost unchanged.
• The extra release files are alternatives and optional material. The new Windows ZIP is about 1.80 GB and the Ubuntu installer about 1.69 GB; choose the format appropriate for your system. The 141 MB technical ZIP contains matching source, licenses and reports for all three platforms and is optional for installation. You do not need to download every asset.

Verified on this exact source: 657 tests on Node 22; Node 24 and browser/setup/settings checks passed. Native Windows installer/application/speech and Apple Silicon ZIP/application/speech checks passed. Local Linux passed 22 package checks, full-context 9B Vulkan conversation with Bambi Mix, a spoken restart with networking disconnected, and a complete fresh model/voice download with CPU conversation. Ubuntu 24.04 passed 22 installed application checks, then another 22 checks after installation of the actual Debian package, with the Chromium sandbox active. See the optional reports for test conditions and limits.

Alpha scope: Full-model Windows 11 GPU/CPU/offline, Apple Silicon Metal/CPU/offline and Ubuntu AI/speech conversations, received-download security, physical microphone/listening and sleep/wake need recipient testing. Windows is unsigned; Mac is ad-hoc signed and not notarized. Keep normal OS security controls enabled. Intel Macs, Linux ARM, devices, online media and automatic updates are not qualified in this release.

Getting started · Settings & presets · Optional matching source, notices and reports · Checksums

Report your app version, OS/hardware, model/backend, steps and result; omit private conversations, recordings and credentials.

Bambi Local alpha.29 · Apple Silicon Mac tester

Download for Apple Silicon Mac · Windows / Linux alpha.23

Mac tester: Apple Silicon (M series), macOS 14 or later. Extract the ZIP and copy Bambi Local to Applications. This build has an ad-hoc signature and is not notarized. Keep normal macOS security controls enabled; report a blocked download or launch.

Try the offline sample, player controls, Studio import/preview/export, and Extras → Settings & presets. Setup explicitly downloads or imports local model and speech data; allow 15 GB free disk space. No models download automatically, and Setup checks memory before allowing model preparation.

CI verified: 26 checks on the finished ZIP cover packaging, player, Studio, settings and process cleanup. Native speech synthesis/recognition and Metal device discovery with CPU fallback planning passed; 516 source tests passed on Node 22 and 24.

Still needs Mac testers: full reference-model CPU/Metal conversations, model setup and disconnected-network restart; downloaded-app/Gatekeeper, audible output, microphone permission/capture/revocation, and sleep/wake. This build remains built-unqualified. Report the app version, Mac/macOS, selected model/backend, steps and result; omit private content.

Tester guide (optional technical ZIP with the guide, matching source, notices and test reports) · Checksums

Windows and Linux downloads remain on alpha.23. Downloads are available without signing in.

Bambi Local alpha.23

Download Windows · Download Linux

Download one file for your system. Open the app, save your access choices and choose a model in setup. The reference 9B model plus speech data needs about 6.05 GB; allow 15 GB free space.

Includes local AI and voice, model-file/folder import, and Extras → Settings & presets for export, import and shared presets.

Alpha testing: Windows 11 full-model GPU/CPU/offline, physical audio and sleep/wake checks remain open. macOS is not available yet. Keep normal OS security controls enabled.

Start guide · Settings guide

The optional technical ZIP contains source, notices and test reports. Most users only need the installer/AppImage. Downloads are available without signing in.