Atomic-Chat version history

- Added speculative decoding stats to the UI. - Improved Windows DLL bundling and backend preferences. - Enabled resumable model downloads with exponential backoff. - Enhanced MLX load safety and speed with proper classification and locks.

- Added DFlash support for llama.cpp, enabling faster token generation. - Integrated OpenClaude for seamless local API server launches. - Introduced new startup toggles for improved workflow customization. - Fixed various crashes and improved reliability across platforms.

- Added support for the Laguna XS 2.1 33B model and TurboQuant llama.cpp for improved performance. - Integrated VS Code, JetBrains, and Xcode for seamless local server connections. - Introduced auto-detection for models from Ollama, LM Studio, HF cache, and Unsloth. - Fixed custom provider model listing and improved macOS window dragging functionality.

## 🚀 New features - Added support for the Laguna XS 2.1 33B model by Poolside via TurboQuant llama.cpp — stronger results on multilingual coding and terminal-style tasks. - [TurboQuant](https://github.com/TheTom/llama-cpp-turboquant) support for llama.cpp on Windows and Linux gives you a bigger context window with lower memory usage, great for running agents on local models. - VS Code, JetBrains and Xcode can now connect straight to your local server, with Base URL, key and model copied over automatically. - We added our own quantized GGUF and MLX versions of Atomic Chat models to the catalog. - New integration lets you detect, configure, and launch Zed (the fast Rust-based editor with a native AI agent panel) and Poolside CLI (a terminal agent built on the open ACP protocol), both pointed at your local server. - Atomic Chat now auto-detects models you already have downloaded from Ollama, LM Studio, HF cache and Unsloth, so you can launch them instead of downloading them again. - Ollama is now a built-in provider, so cloud and local Ollama models both show up and launch right away. ## 🔧 Improvements & fixes - The inference server port is now configurable instead of h

## 🚀 New features - [TurboQuant](https://github.com/TheTom/llama-cpp-turboquant) support for llama.cpp on Windows and Linux gives you a bigger context window with lower memory usage, great for running agents on local models. - VS Code, JetBrains and Xcode can now connect straight to your local server, with Base URL, key and model copied over automatically. - We added our own quantized GGUF and MLX versions of Atomic Chat models to the catalog. - New integration lets you detect, configure, and launch Zed (the fast Rust-based editor with a native AI agent panel) and Poolside CLI (a terminal agent built on the open ACP protocol), both pointed at your local server. - Atomic Chat now auto-detects models you already have downloaded from Ollama, LM Studio, HF cache and Unsloth, so you can launch them instead of downloading them again. - Ollama is now a built-in provider, so cloud and local Ollama models both show up and launch right away. ## 🔧 Improvements & fixes - The inference server port is now configurable instead of hardcoded to 1337, and Atomic Chat falls back automatically if that port is already taken by another app ([#115](https://github.com/AtomicBot-ai/Atomic-Cha

## 🚀 New features - Audio input for omni and MLX models — record and play audio right in the chat - The app now remembers window size and position across restarts - Download a model during onboarding and you'll jump straight into chat, with progress still visible - Smarter sampling: recommended defaults for Gemma 4 QAT models, while keeping any values you've customized - MLX models now have a proper Download section on the Hub (name, MLX badge, file size, Download / New Chat) ## 🔧 Improvements & fixes - Fixed CUDA and Vulkan auto-detection on Windows — the right GPU backend now gets picked and downloaded automatically - More reliable backend downloads overall (shared / NAT / VPN networks, fewer GitHub rate-limit failures, offline fallback preserved) - Better Metal GPU recovery: clearer out-of-memory messages and auto-reload after compute errors - Fixed system prompt being dropped in regular local-model chats - Clickable links in the model-policy error banner - Quieter logs (no more misleading CORS warnings for non-browser clients) ## 🙏 Contributors Thanks to [@Vect0rM](https://github.com/Vect0rM), [@Albert-Atomic](https://github.com/Albert-Atomic) for their con

## 🚀 New Features - Add Cline CLI and MiMo Code coding agents to the Integrations page for one-click running - Support WebP image attachments - Add drag-and-drop for projects in the sidebar - Add a configurable maximum image size with automatic downscaling when images exceed the limit - Add Korean (한국어) language support ## 🔧 Improvements & Fixes - Migrate macOS autostart from LaunchAgent to AppleScript, with a one-time migration to preserve existing autostart preferences - Add tiered backend fallback when a pinned backend is unavailable, returns a 404, or is unreachable: newest installed same-type backend → newest published family tag (download) → any installed backend as a last-resort CPU degrade, with corrected pins persisted for tiers 1–2; also clean up orphaned/incomplete backend installs automatically - Resolve the latest published CUDA minor version dynamically in Windows build scripts and handle GitHub API rate limits gracefully, preventing 404s - Map local model-server transport errors to 503 with a structured OpenAI-style error envelope (including Retry-After); wrap local-engine errors with typed codes (e.g. context_length_exceeded, insufficient_memory) - Va

## 🚀 New Features - Added a "Download a model" shortcut to the model dropdown for quicker model downloads from the selector. - On macOS, closing the window now sends the app to the tray instead of quitting so the app keeps running in the background. - Model loading now shows clearer, backend-aware errors for unsupported or corrupt models so you get actionable diagnostics when a model fails to load. ## 🔧 Improvements & Fixes - Search moved to the bottom of the sidebar navigation so the Search item lives at the end of the nav list. - Repair invalid tool-call JSON emitted by coder models instead of crashing, preventing interruptions when models produce malformed tool-call output. - Sidebar tree view polish: tightened the project chevron spacing and aligned per-conversation overflow ("...") buttons so tree items look consistent and overflow actions appear in the right place. - Windows maximize fix: caption buttons now fill the very top edge when the window is maximized so close/minimize/maximize are easier to hit. - macOS agent and backend fixes: prevented the upstream auto-upgrade from deleting unrelated turboquant backends, added recovery to restore missing turboquant ba

## 🚀 New Features - Global sampling settings — sampling parameters are now a single persisted global store (Zustand) instead of per-assistant, so sampling behaves consistently across the app and legacy on-disk data is preserved. - Pause/resume model downloads — resumable GGUF downloads can be paused and resumed from the global Downloads popover with new UI controls and resume-state handling for better download control. ## 🔧 Improvements & Fixes - Clearer model loading feedback — previously persistent "Failed to load" toasts are cleared on successful load and engine error codes are mapped to actionable titles/descriptions, with updated English and Russian locale messages. - Windows GGUF validation fix — model path validation was enhanced to avoid issues with split GGUF files on Windows, improving model loading reliability. - Download UI and robustness tweaks — download button keeps consistent height during download states and paused downloads no longer trigger error notifications or cleanup. ## 🙏 Contributors Thanks to [@Vect0rM](https://github.com/Vect0rM) for their contributions to this release!

## 🚀 New Features - Add one-click Kilo, Pi, Goose and OpenHands on the integration page to quickly start coding agents - Enable "Launch at startup" by default for all users - Render inline image attachments directly inside chat bubbles so images appear in-thread ## 🔧 Improvements & Fixes - Prevent non-MTP models from being bricked by gating the global llamacpp-upstream MTP flag on a per-model capability at load time. - 5+ stability and UI fixes ## 🙏 Contributors Thanks to [@Vect0rM](https://github.com/Vect0rM) and [@danyurkin](https://github.com/danyurkin) for their contributions to this release!

## 🚀 New Features - **Launch at startup** — new toggle in Settings → General to start Atomic Chat automatically when your system boots (macOS, Windows, Linux) ## 🔧 Improvements & Fixes - **macOS recommended model now loads out of the box** — the default llama.cpp engine is now vanilla upstream, so the Gemma vision model no longer crashes on load. TurboQuant stays available as an opt-in - **Readable model-load errors** — failed loads now show the real reason instead of `[object Object]` - **Clearer Trusted Hosts setup** — reaching the local API server over LAN now tells you exactly which address to add (use `*` to allow all) - **"Launch With" list** — added Kilo Code, sorted by popularity. See the full list in the [README](https://github.com/AtomicBot-ai/Atomic-Chat#launch-with) ## 🙏 Contributors Thanks to [@Vect0rM](https://github.com/Vect0rM), [@yanalialiuk](https://github.com/yanalialiuk)!

# Atomic Chat v1.1.104 ## 🚀 New Features - Gemma 4 12B now runs in Atomic Chat — Google's latest multimodal model, working on Mac (MLX) and Windows/Linux (llama.cpp). - Added Google's Gemma 4 QAT models support — near-original quality using up to ~72% less memory, with sizes from 2B up to 31B. ## 🔧 Improvements & Fixes - Updated llama.cpp to the latest build, enabling faster Gemma 4 generation (MTP speculative decoding). - Fixed a Windows bug where a fresh install couldn't save its settings, which could break first-run onboarding. - Improved CUDA backend installation on Windows to fix cases where it failed to set up after updating. - Future-proofed Windows GPU acceleration so it won't silently drop to CPU when the CUDA build version changes. - Fixed the OpenClaw config being rejected so one-click Run works again without manual edits. ## 🙏 Contributors Thanks to [@Vect0rM](https://github.com/Vect0rM), [@Albert-Atomic](https://github.com/Albert-Atomic) for their contributions to this release!