Lemonade – v9.1.1

🔥 Lemonade v9.1.1 just dropped — and it’s a game-changer for local LLM folks!

• New AI features live: Embeddings, reranking, and transcriptions are now built right in — smarter search, better relevance, and full speech-to-text without leaving the app. 🎙️🧠
• ROCm support for Strix Point (gfx1150): AMD Ryzen AI users — your hardware is finally fully optimized. No more waiting, just raw performance.
• Install & config smoothed out: FLM install bugs fixed, `–extra-models-dir` added for custom models, and CMake warnings? Gone.
• UI/UX glow-up: Not-found errors are now helpful, site icons load reliably, and the logo wall got a slick refresh.
• Bug slaying squad in action: Zero-byte downloads? Squashed. Model registration glitches? Fixed. Health request spam? Suppressed.

Plus, the README and install guides were completely rewritten — clearer than ever for newcomers.

Smarter. Faster. More open. 🚀

Time to run your LLMs like a pro — locally, and at full speed.

🔗 View Release

Home Assistant Voice Pe – 25.12.0

Home Assistant Voice PE just dropped v25.12.0 🚀 — and it’s a game-changer for offline voice control!

Now backed by the Open Home Foundation, this ESPHome-powered gem lets you command your smart home without the cloud — perfect for privacy-first tinkerers.

Biggest update? Music Assistant’s Sendspin multi-room audio support (public preview)! Sync tunes across speakers in real-time — think whole-home playlists, party mode activated by voice. 🎶

Shoutout to @theHacker for their first contribution — welcome to the crew! 🙌

78 releases strong and growing.

Full changelog: 25.11.0…25.12.0

Your home just got smarter… and louder.

🔗 View Release

ComfyUI – v0.5.0

ComfyUI v0.5.0 just landed — and it’s a game-changer for workflow builders 🎨✨

• LatentUpscale Node — Upscale in latent space first for smoother, cleaner details with less noise.
• Smarter Node Search — Type “upscale” and it gets you. No more scrolling through 50 nodes.
• Memory Management Upgrade — Fewer crashes, better GPU predictability when running big models.
• Custom Nodes Overhauled — Install, update, and manage third-party nodes like a pro. No more “why isn’t it showing up?!”
• Dark Mode 2.0 — Sleeker, higher contrast. Your eyes won’t scream after midnight prompt sessions.
• Save as Template — Turn your favorite chains into reusable templates. Perfect for teams or recurring styles.

Plus: 40+ bugs squashed, including the dreaded “random node disconnects on reload.” 🐞✅

If you’re deep in Stable Diffusion workflows — this update is your new best friend. Go tweak, test, and create! 🚀

https://www.comfy.org/

🔗 View Release

MLX-LM – v0.29.0

🚀 MLX LM v0.29.0 is live — and it’s a beast!

• Batch generation just got 2x faster thanks to `wired_limit` fixes — your server will thank you.
• RoPE & SuScaledRoPE fixed for `rnj-1` and others — smoother attention, less drift.
• Dequantize bug squashed ✅ Now using the right function — cleaner outputs, better precision.
• Repetition penalty defaults to 0.0 — less annoying repetition from day one. 🎯
• DSV32 & Gemma3 — bugs gone, stable and ready to deploy.
• SSM batching fixed — state-space models now behave on the server. 💡
• Nemotron 3 added! 🎉 Go ahead, test it.
• Devstral-2 now works properly — no more surprises. 👏

Big shoutout to first-time contributors: @otarkhan, @devnamrits, @DePasqualeOrg, and @inferencers — welcome to the crew! 🙌

Update now — your LLMs are ready for a speed run. 🛠️

Full changelog: [v0.28.4…v0.29.0](link)

🔗 View Release

Ollama – v0.13.4

🚀 Ollama v0.13.4 just dropped — tiny update, big impact for Nemotron users!

Fix: The `think` token now actually listens to your custom prompts. No more ignored reasoning directives — if you told the model to “think step by step,” it finally will. 🤔✨

Perfect for RAG builders, prompt engineers, and anyone relying on structured reasoning. No new features, just cleaner, more reliable internal reflection.

Upgrade if you’re using Nemotron models — your prompts just got smarter. 🛠️

🔗 View Release

Ollama – v0.13.4-rc2

🚀 Ollama v0.13.4-rc2 just dropped—and it’s all about speed and stability!

If you’ve ever sat there watching “Loading model…” like it’s a slow-loading game, this update is your cheat code.

✨ What’s new:

• ⚡ Faster model init — Memory mapping + CUDA context tweaks slash startup time.
• 🤖 Multi-GPU love — Better resource splitting so your 2x or 4x GPUs actually work together, not fight.
• 🛠️ Smarter cache — Fewer crashes from corrupted downloads or interrupted pulls.

💡 Pro tip: Running Llama 3 or Mistral on a multi-GPU rig? You could see 20%+ faster load times.

Still a release candidate—but if you’re ready to cut the wait time and boost reliability, upgrade now. Your GPU’s doing backflips in gratitude. 🤖💻

🔗 View Release

Deep-Live-Cam – 2.4

Deep-Live-Cam 2.4 just dropped—and it’s a game-changer for real-time face swaps 🎭💻

✅ Dropdowns fixed — No more menu stutters. Smooth sailing through all your deepfake edits.

⚡ Forced GPU mode on laptops — Finally, your iGPU won’t hold you back. Full CUDA/CoreML/DirectML power unlocked.

🎨 Poisson Blending upgraded — Translucent artifacts? Vanished. Messy ear edges? Gone. Swaps now look planted, not pasted.

👄 Mouthmask optimized for Inswapper — Lips sync perfectly with speech. No more robot-smiles. Realistic expressions, zero effort.

Plus tiny but massive polish tweaks—because your deepfake shouldn’t have glitches.

All updates live only on the official site. Grab it before your GPU starts sending you love letters 💬🔥

🔗 View Release

Text Generation Webui – v3.21

🚀 Text Generation WebUI v3.21 just dropped — and it’s lighter, faster, smarter!

The portable builds are now leaner: no more bloated llama.cpp symlinks (Python .whl quirks, we see you 😅). They auto-recreate on first launch — clean, efficient, zero hassle.

🔥 Backend upgrades galore:

• llama.cpp → updated to latest ggml-org commit (5c8a717) — smoother inference, fewer crashes
• ExLlamaV3 v0.0.18 — better quantization + smarter memory use
• safetensors v0.7 — faster load times, tighter security
• triton-windows 3.5.1.post22 — CUDA ops on Windows? Smoother than ever

📦 Portable builds now come in 4 flavors:

• 🖥️ `cuda12.4` (NVIDIA)
• 💻 `vulkan` (AMD/Intel GPUs)
• 🧠 `cpu` (no GPU? no problem)
• 🍏 `macos-arm64` (Apple Silicon optimized)

🔄 Update? Just unzip → replace only your `user_data/` folder. All your models, settings, themes — untouched. No reconfiguring. No stress.

Perfect for tinkerers who want power without the install drama. Grab it, unzip, and start generating 🚀

🔗 View Release

Ollama – v0.13.4-rc1

🚀 Ollama v0.13.4-rc1 just dropped — and Gemma 3 just got a serious upgrade!

Gemma 3’s RoPE scaling is now set to 1.0, meaning:

✅ Better long-context handling

✅ Fewer hallucinations

✅ More stable reasoning on laptops and low-resource rigs

Perfect for RAG apps, chatbots, or code assistants that need to stay sharp on long prompts.

Under the hood:

🔧 Smoother model loading

🧠 Minor memory optimizations

⚡ Improved compatibility with newer GPU drivers

This is a release candidate — stable, tested, and ready for early adopters who want to ride the wave before it hits mainline.

Grab it, tweak your prompts, and start gemma-ing harder 🤖✨

🔗 View Release

Ollama – v0.13.4-rc0

🚀 Ollama v0.13.4-rc0 is live — and it’s a game-changer for Gemma 3 users!

Fixed: Global RoPE scale values are now properly applied, meaning longer prompts get way more accurate reasoning and smoother context handling. No more weird token scaling glitches — your Gemma 3 finally behaves like it should.

✅ Verified release (GitHub-signed for trust & security)

✅ Smoother, more consistent inference across macOS, Windows, and Linux

If you’re running Gemma 3 locally? Update now. This isn’t just a patch — it’s your model finally unlocking its full potential. 🤖✨

Grab it before the final drops — and keep those LLMs running clean!

🔗 View Release