Vesta AI Explorer for macOS - macOS Tahoe 26

vesta-mac.pages.dev

Install with Homebrew (recommended)

brew tap scouzi1966/afm
brew install --cask scouzi1966/afm/vesta-mac

Direct Download

Channel	Version	Download	Date
Stable	0.9.7	Vesta-0.9.7.dmg	2026-05-08
Nightly	Vesta-next	Vesta-next.dmg	Vesta-next

Nightly via Homebrew

brew install --cask scouzi1966/afm/vesta-mac-next

Vesta is a multi-backend AI chat application for macOS that runs models locally on Apple Silicon. It supports five AI backends simultaneously -- Apple Intelligence, MLX, llama.cpp, HuggingFace Inference API, and OpenAI-compatible API servers -- with runtime switching, vision understanding, text-to-speech, speech-to-text, image generation, video generation, and a full MCP server for programmatic control.

Demo

Click to watch the full demo video

Note that there is a known bug in Qwen3-VL models where it will loop indefinitely in some cases and never complete. In this case stop the generation and tweak your prompt. This is a model behavior, not the app. Reference: https://github.com/QwenLM/Qwen3-VL/issues/1611#issuecomment-3639175711

---

Vesta 0.9.7 (May 8, 2026)

MLX model support sync with maclocal-api and release pipeline hardening.

Changes

• Qwen 3.6 support -- imported maclocal-api architecture mapping for Qwen 3.6 MLX models.
• Gemma 4 support -- aligned Gemma 4 model handling and chat template fallback behavior with maclocal-api.
• Crash fix -- corrected Gemma 4 VL attention reshape handling for prompt generation.
• Release publishing -- rebuilt from scratch in Release, notarized, stapled, and published as a signed DMG.

---

Vesta 0.9.6 (Mar 4, 2026)

Stability and release pipeline improvements.

Changes

• Stacked download button -- version and date displayed vertically on the website download button
• Publish date on stable button -- stable download button now shows the release date
• Expanded test suite -- automated test pipeline expanded to 4 stages with full coverage
• Pre-test validation -- release skill checks for running Vesta instance before tests

---

Vesta 0.9.5 (Feb 11, 2026)

94 commits, 2 months of development since v0.9.0

This is a major release that transforms Vesta from a two-backend chat app into a full multi-modal AI platform with five backends, MCP integration, and media generation capabilities. Explore the world of AI beyond ChatGPT!

New in 0.9.5

Multi-Backend Architecture

• HuggingFace Explorer -- Leverage the HuggingFace Pro subscription and inference providers partners for ImageGen, Videogen, Image edit, Whisper etc. I am not affiliated with HuggingFace but the pro sub is worth it.
• 5 simultaneous backends -- Apple Intelligence, MLX, llama.cpp, HuggingFace Explorer, and External AI (OpenAI-compatible) all active at once (Single chat routing)
• Runtime backend switching -- switch between backends without restarting the app
• Per-backend settings -- each backend has its own generation parameters, model selection, and configuration
• HuggingFace Browser -- Direct browse and download models from HuggingFace in-app

MCP Server (Model Context Protocol) - Agentic Sidekick!

• Full MCP server running on TCP loopback with token-based authentication
• Agentic Sidekick Vesta detects Claude Code when you enable MCP - Claude Code acts as an agent with a NLI (Natural Language Interface) to the app. Ask Claude to set things up, have a conversation with any other model!
• 41+ tools -- backend management, chat, model download/load/unload, vision analysis, settings, conversation history search, diagnostics, UI navigation
• 6 resources -- app state, models, conversation, settings, logs, system info
• 7 prompts -- guides for Vesta, MLX, llama.cpp, HuggingFace, and common workflows
• AI Sidekick -- Claude Code integration for programmatic Vesta control
• Conversation history search -- full-text search and read-only SQL queries against the SQLite message database

HuggingFace Explorer (New Backend)

• Cloud inference via 16+ providers (Cerebras, Groq, Together, Fireworks, SambaNova, Nebius, Replicate, and more)
• Text-to-image generation -- FLUX.1 Schnell/Dev, FLUX.2, Stable Diffusion 3.x, SDXL with configurable size, guidance, and steps
• Image editing -- instruction-based editing with FLUX Kontext and FLUX.2 via Replicate
• Video generation -- Wan 2.2 T2V (text-to-video) with async polling
• Speech-to-text transcription -- OpenAI Whisper models via HuggingFace Inference API with 14 languages
• Vision/VLM -- Qwen2.5-VL, Qwen3-VL and other vision-language models via cloud
• Model browser -- search and discover models from HuggingFace Hub with download counts, likes, and gated model detection
• Reasoning display -- chain-of-thought rendering for models that emit <think> tags (DeepSeek R1, QwQ, etc.)
• 7-tab settings panel -- Chat, Vision, Image, Edit, Transcribe, Video, Settings

Text-to-Speech (TTS)

• Kokoro (82M) -- 46+ voices across 10+ languages, fast and high-quality
• Marvis (100M/250M) -- conversational TTS with voice cloning support via reference audio
• Models download on first use from HuggingFace

Speech-to-Text (STT) -- WhisperKit

• On-device transcription via WhisperKit CoreML -- zero network required
• 6 model sizes -- Tiny (39M) through Large V3 (1.5B) and Large V3 Turbo (809M)
• 28+ languages with auto-detect
• Per-segment timing and speed ratio reporting

Jinja Template Support (minja)

• Full Jinja2 template parsing for GGUF models via llama.cpp's minja library
• Correctly renders chat templates embedded in model metadata
• Falls back to llama_chat_apply_template() for non-Jinja templates

GGUF Model Browser

• Search and browse GGUF models from HuggingFace Hub
• Capability badges: Vision, Tool Use, Reasoning, Coding, Math, Multilingual
• Split/multipart file detection
• Automatic mmproj detection for vision models

MLX Improvements

• Qwen3-VL M-RoPE patch -- +81% performance improvement for vision inference (auto-applied via script)
• KV cache controls exposed -- max KV size, quantization bits, prefill step size
• MLX benchmark tool (mlx-bench) for standalone performance testing
• Wired memory set to 90% of GPU recommended working set for large model performance

Improvements

• Per-message metrics -- token count and tokens/sec stored with each message in the database
• Green parameter labels -- visual indicator when a generation parameter matches the model's configured default
• Binary voice format -- Kokoro voice files converted from JSON to binary (144 MB down to 27 MB)
• Conversation history view with backend filtering and pagination
• One-command build -- build-from-scratch.sh handles submodules, patching, llama.cpp library build, and Xcode build
• Distribution pipeline -- automated DMG creation, notarization, and GitHub release via build-vesta-mac-dist.sh
• Automated testing framework -- 44+ UI tests via MCP-based test runner

Bug Fixes

• Fix AVKit VideoPlayer crash during SwiftUI transitions (disabled transition animations)
• Fix download progress stuck at 0% and crash in llama.cpp streaming
• Fix TTS mode hijacking text generation when both TTS and LLM models are loaded
• Fix O(N^2) reasoning parser performance
• Fix reasoning parser stripMarkers bug and chat history contamination
• Fix streaming throttle not kicking in when content scrolls off-screen
• Fix llama.cpp default context size (2048 changed to 16384 for Qwen3-VL)
• Fix Continuity Camera Swift 6 concurrency crash in Release builds (Objective-C workaround)
• Fix NSHostingView constraint crash in MLX settings window (non-observing wrapper)
• Fix MoE warmup crash for models with 32+ experts (reduced warmup batch size)
• Fix mxfp4 MoE Metal shader crash (skip warmup for mxfp4 models)
• Fix ESpeakNG unsealed contents causing notarization failure
• Fix GGUF vision model image handling and model deduplication

---

Vesta 0.9.0 (Dec 10, 2025)

New in 0.9.0

• Vision capabilities with Qwen3-VL model (describe images, analyze screenshots)
• Continuity Camera input (capture photos directly from iPhone/iPad)
• Code syntax highlighting for 20+ programming languages
• Edit responses inline after generation
• HTML preview for rendered content
• Enhanced LaTeX math rendering in blockquotes
• Improved rendering engine with real-time code block highlighting

---

Features

• Apple Intelligence -- on-device AI via Foundation Models framework (always available)
• MLX Backend -- Apple Silicon optimized inference with mlx-swift (Qwen3-VL vision models)
• llama.cpp Backend -- GGUF model support with full Metal GPU acceleration and Jinja templates
• HuggingFace Explorer -- cloud inference, image/video generation, transcription via 16+ providers
• External AI -- connect to any OpenAI-compatible API server (LM Studio, Ollama, etc.)
• Vision -- image understanding via Qwen3-VL (MLX, llama.cpp, or HuggingFace)
• Text-to-Speech -- Kokoro, Marvis (with voice cloning), and Orpheus TTS engines
• Speech-to-Text -- WhisperKit on-device transcription (Tiny through Large V3)
• MCP Server -- 41+ tools for programmatic control, model management, and AI Sidekick integration
• GitHub Flavored Markdown -- tables, task lists, strikethrough via remark/rehype pipeline
• LaTeX Math -- inline and block math rendering with KaTeX
• Code Highlighting -- 20+ languages with real-time streaming highlight
• Liquid Glass UI -- native macOS Tahoe design
• App Sandbox -- Developer ID signed and Apple notarized

Verify

# Check SHA256
shasum -a 256 ~/Downloads/Vesta-*.dmg

# Check code signature
codesign --verify --deep --strict /Applications/Vesta.app

# Check notarization
spctl --assess --type execute /Applications/Vesta.app

Requirements

• macOS 26.0 (Tahoe) or later
• Apple Silicon Mac (M1/M2/M3/M4)
• Microphone access for voice input and STT
• Internet access for HuggingFace backend and model downloads (on-device backends work offline after model download)

Security & Privacy

• Signed with Developer ID Application: Soprano Technologies Inc.
• Notarized by Apple
• App Sandbox enabled
• On-device backends (Apple Intelligence, MLX, llama.cpp) process everything locally -- no data sent to servers
• HuggingFace and External AI backends require network access for inference
• API tokens stored in macOS Keychain

Related

• Source Code: https://github.com/scouzi1966/vesta-mac
• Distribution: https://github.com/scouzi1966/vesta-mac-dist (this repo)
• Website: https://vesta-mac.pages.dev
• CLI Alternative: https://github.com/scouzi1966/maclocal-api

Support

• Report an Issue
• Request a Feature
• Browse Issues

License

(c) 2025-2026 Soprano Technologies Inc. All rights reserved.

Built With

• Apple Intelligence -- Foundation Models framework
• MLX -- mlx-swift + mlx-swift-lm for Apple Silicon inference
• llama.cpp -- GGUF inference with Metal acceleration
• WhisperKit -- CoreML-based Whisper speech-to-text
• mlx-audio -- Kokoro/Marvis/Orpheus TTS
• SwiftUI -- native macOS interface
• KaTeX -- math rendering
• highlight.js -- code syntax highlighting
• remark/rehype -- markdown processing pipeline

---

Built with automated distribution pipeline | Notarized and code-signed | Apple Silicon native

Star History