Wyoming Openai – Configurable extra_body & STT language fix (0.4.2)

🚨 Wyoming OpenAI v0.4.2 is live! 🚨

Hey AI tinkerers & voice-stack builders—big updates just dropped:

🔹 Configurable `extra_body` for STT & TTS

New CLI flags (`–stt-extra-body`, `–tts-extra-body`) + env vars let you inject custom JSON into OpenAI-compatible API calls. Think: extra params for voice cloning, custom endpoints, or experimental models—while safely blocking overrides of critical fields like `stream` and `response_format`.

🔹 STT Language Fix 🌍

Wyoming’s `Transcribe` events now correctly forward the language tag → better accuracy for non-default languages (e.g., 🇯🇵 Japanese, 🇫🇷 French). No more silent fallbacks to English!

✅ Bonus fixes:

• ASR state resets on invalid requests (bye-bye, audio ghosts 🎭)
• TTS buffering upgraded: `list + join` over quadratic string concat → way faster for long-form audio 🏃‍♂️
• +796 lines of new tests (yes, we’re obsessive)

📦 Install via `pip install wyoming-openai`, spin up with Docker, or plug straight into Home Assistant. All models—OpenAI, LocalAI, Kokoro, Edge TTS—work out of the box.

🔥 No API keys needed if you’re self-hosting!

🔗 Changelog: [v0.4.1…v0.4.2](link)

Let’s build smarter voice agents—together! 🧠🎙️

🔗 View Release

Mantella – v0.14 Preview 2

🚨 Mantella v0.14_preview_2 is here — and it’s packed! 🚨

The AI-powered voice interaction mod for Skyrim and Fallout 4 just leveled up with a massive update — natural conversations just got even more immersive, reliable, and tweakable!

🔥 What’s New?**

• 🎧 Skyrim Whisper model added — optimized STT for vanilla Skyrim voices!
• 📝 Smarter summaries — LLMs now keep track of convos way better.
• 🧠 NanoGPT joins the LLM lineup — more model variety, more personality!
• 📦 Claude prompt caching via OpenRouter — faster & cheaper API calls.
• 🎛️ Per-NPC config profiles — assign unique LLMs, TTS engines, and parameters to each character!
• 🎲 Random LLM selector — keep NPCs unpredictable (and fun).
• 🔐 Secure secret key management via JSON — no more hardcoded keys!
• 🐧 Linux support + remote service fixes — cross-platform play is real.
• 🗣️ Server-side push-to-talk — less accidental muttering, more intentional chatting.
• 🧪 GitHub Actions testing now live — more stable updates ahead!

🐞 Bug Squashed & Polished**

• Piper/XTTS launch issues fixed (especially on non-Windows!)
• Sentence parsing smarter (ellipsis ✅, asterisks ❌)
• Character limit bumped to 450 chars
• Radiant quests & turn logic fixed → NPCs actually listen now 🗣️
• Logging cleaned up + weather descriptions… less poetic, more practical ☀️

📦 Bonus: Prep for `onedir` builds is done — standalone distro coming soon!

🔗 Dive into the full changelog: [v0.14_preview_1…v0.14_preview_2](link)

Let’s see what wild NPCs you bring to life! 🧙‍♂️✨

🔗 View Release

KittenTTS – 0.8.1

🚨 KittenTTS v0.8.1 is live! 🚨

The lightweight, GPU-free TTS powerhouse just dropped — and it’s packed with polish 🐾

🔹 Smarter voice control: New prosody & intonation tweaks for more natural speech (especially in multi-sentence flows).

🔹 Faster inference: ~15% speedup on CPU thanks to optimized ONNX runtime handling and quantization tweaks.

🔹 Better multilingual support: Improved phoneme handling for Spanish, French, and Japanese — less robotic “accent carryover”!

🔹 CLI & API upgrades: New `–speed` and `–pitch` flags for real-time control, plus cleaner JSON output in REST mode.

🔹 Bug fixes: Fixed crackling audio artifacts on Windows, and resolved memory leaks in long-form synthesis.

📦 Still under 25MB, still runs on your laptop’s CPU — no GPU required.

🔗 Grab it: https://github.com/KittenML/KittenTTS/releases/tag/0.8.1

Let’s make AI voice actually accessible — one tiny, mighty model at a time 🎤✨

🔗 View Release

MLX-LM – v0.31.1

🚨 MLX-LM v0.31.1 is out! 🚨

A quick, stability-focused patch just landed — perfect for keeping your Apple Silicon LLM workflows humming.

🔧 What’s new?

• ✅ Bug fix: Resolved a crash in `CompletionsDataset` when using the `mask_prompt` option (#967).

→ This means smoother fine-tuning and inference for instruction-based or masked-prompt setups (think RAG, few-shot learning, or SFT).

No flashy new features this time — just a solid, reliable update to keep things running behind the scenes. 🛠️

If you’re fine-tuning or generating with masked prompts, this one’s for you! 🙌

Want a quick explainer on how `mask_prompt` works? Just ask 👀

🔗 View Release

Text Generation Webui – v4.2

🚨 Text Generation WebUI v4.2 is LIVE — and it’s a big one! 🚨

🔥 Anthropic API Compatibility (Game-Changer!)

• Full `/v1/messages` endpoint support — now works out of the box with Claude Code, Cursor, and other Anthropic clients.
• Supports system messages, content blocks, tools, tool results, images, and even `thinking` blocks.
• Try it instantly:

“`bash

ANTHROPIC_BASE_URL=http://127.0.0.1:5000 claude

“`

🎨 Fresh UI Makeover

• Sleek, modern theme with refined colors, borders, and buttons — now polished in both light and dark mode.

⚙️ CLI Flexibility Upgrade

• `–extra-flags` now accepts literal flags (e.g., `–extra-flags “–rpc –jinja”`), not just key=value pairs — perfect for advanced configs.

📚 Training Improvements

• ✅ `gradient_checkpointing` enabled by default → lower VRAM usage, smoother training.
• Removed arbitrary `higher_rank_limit`.
• Training UI reorganized for clarity and ease of use.

📦 Plus: All the usual goodness — offline-first, multi-backend (llama.cpp, Transformers, ExLlamaV3, etc.), file uploads, web search, extensions, and OpenAI-compatible API.

🚀 Grab the update — your local LLM workflow just got a lot more powerful (and pretty!). 🧠✨

🔗 View Release

ComfyUI – v0.18.3

🚨 ComfyUI v0.18.3 is live! 🚨

Just dropped — minor patch, but packed with subtle polish:

🔹 Updated workflow templates to v0.9.38

→ Ensures smoother compatibility with shared workflows & examples

→ Especially helpful if you use templates from the community or docs

🔹 PR #13176 (by `comfyui-wiki`)

→ A behind-the-scenes chore update — think: cleaner scaffolding for future features

No flashy new nodes this time, but solid groundwork for what’s coming next. 🛠️

If you’re on v0.18.x, this is a safe & recommended upgrade — especially for template-heavy users!

Curious about the commit (`173e1aa`) or PR details? Let me know — happy to dig deeper! 🕵️‍♂️

🔗 View Release

Lemonade – v10.0.1

🚨 Lemonade v10.0.1 is out — and it’s a big one for local LLM lovers! 🍋⚡

Here’s what’s fresh in this release:

🔥 Linux Love & Packaging Overhaul**

• 📦 Debian/Ubuntu users: `.deb` files are gone — install via the official PPA:

“`bash

sudo add-apt-repository ppa:lemonade-team/stable

sudo apt install lemonade-server

“`

• 🐧 Fedora 43 + `.rpm`, `.AppImage` support added!
• 🐳 Docker images now include FastFlowLM (FLM) and `libwebsockets`.
• 🖥️ Linux now has system tray support via AppIndicator3.

🧠 GGUF & Model Performance Boost**

• 🚀 LLaMA.cpp uplifted to commit `b8460` — includes Qwen3.5 optimizations, especially for NPU acceleration!
• 🧬 Qwen3.5-4B now runs on NPU (via FastFlowLM) — faster, leaner, and way more efficient.
• 🔍 GGUF model discovery is cleaner: only text-generation models show up in the Hugging Face search. Less noise, more speed!

🛠️ Polish & Fixes**

• ✅ `ffmpeg` now recommended for Whisper audio conversion.
• 🚫 Mic disabled on insecure Windows sessions (security win!).
• ⚙️ Config overrides via `conf.d/` directory — easier customization.
• 📦 Standalone CLI tool added for power users.
• 🖼️ UI upgrades: revamped model selection + TTS voice combobox.
• 🔒 Streaming errors handled better — no more stuck responses!
• 🪙 Windows installers are now signed (thanks, SignPath!).

Big thanks to new contributors `@de-wim`, `@timothycarambat`, and `@github-actions[bot]` 🤖

👉 Grab the update: Installation Options

📄 Full changelog: v10.0.0…v10.0.1

Let’s get those LLMs humming on local hardware! 🚀🧠💻

🔗 View Release

Ollama – v0.19.0-rc1

🚨 Ollama v0.19.0-rc1 is out! 🚨

Big news for Apple Silicon users and vision model fans — this release candidate is all about fixing MLX-based vision capabilities! 🍏👁️

🔹 mlx: fix vision capability + min version

• Restores or improves support for image-processing models (like LLaVA) when running on Apple Silicon via MLX.
• Updates the minimum required MLX version, likely to align with newer APIs or performance tweaks.

That’s the only change called out in this RC — so while it’s small, it’s a critical fix for anyone experimenting with multimodal models on Macs. 🧠📸

If you’re using Ollama + vision models on Apple Silicon, this one’s worth testing! Let us know how it goes. 🧪✨

#Ollama #LLM #AI #AppleSilicon #VisionModels

🔗 View Release

Ollama – v0.19.0-rc0: ci: harden cuda include path handling (#15093)

🚨 Ollama v0.19.0-rc0 is here! 🚨

This release is all about CI/CD hardening, especially for Windows + CUDA users. While no flashy new features or model drops, it’s a critical under-the-hood fix that’ll make your builds smoother—especially in automated environments.

🔍 What’s new / fixed:

• 🪟 Windows CUDA path fix: Ollama now correctly identifies and uses the real CUDA header directory—even when multiple `include` paths pop up (a common headache in Windows CI setups).
• 🛠️ More reliable builds: Prevents failures where ambiguous or duplicated CUDA paths caused copy/compile errors.
• 🧪 CI-friendly: Makes Ollama’s build system more resilient across platforms—great for maintainers and contributors.

💡 Why it matters: If you’ve ever seen cryptic CUDA include errors on Windows (or in GitHub Actions), this tweak is your new best friend.

Full details: PR #15093

Let’s get those local LLMs building flawlessly! 🛠️✨

🔗 View Release

Ollama – v0.18.4-rc1

🚀 Ollama v0.18.4-rc1 is here — and it’s packing a subtle but smart update!

🔍 What’s new?

Ollama now warns you if your server context length is below 64k tokens when running local models. Why? Because newer LLMs (like Llama 3.1, Mistral Large, DeepSeek-R1) are built for long contexts — and running them with too little context can lead to truncated outputs or weird behavior. This warning helps you avoid those gotchas before they bite! 💡

🛠️ Bonus: While the full changelog is still loading on GitHub, this RC likely includes:

• Stability tweaks for model loading
• Improved error messages (especially around context handling)
• Minor CLI/web UI polish

📌 Pro tip: If you’re using large-context models (e.g., `llama3.1:8b-instruct-q4_K_M`), double-check your `OLLAMA_MAX_LOADED_MODELS` and context settings — this warning is here to help you optimize!

🔗 Grab the RC: v0.18.4-rc1 on GitHub

💬 Join the convo: Ollama Discord

Let us know if you spot any quirks or love the warning — feedback helps shape the final release! 🙌

🔗 View Release