Category: AI

Ollama – v0.13.0-rc0

🚀 Ollama v0.13.0-rc0 just dropped — and it’s packed with power!

Say hello to DeepSeek-V3.1 (aka Deepseek2) — one of the most capable open LLMs out there, now available with a simple `ollama pull deepseek-ai/deepseek-v3.1`.

✨ Why it’s awesome:

• 🚀 MLA (Multi-Layer Attention) is live — cuts memory use, speeds up inference, and keeps reasoning sharp.
• 🛠️ New engine under the hood = smoother runs, fewer crashes, better future-proofing.
• 💥 Run state-of-the-art reasoning on your laptop — no cloud needed.

GGUF? Still supported. API? Still there. CLI? Even better.

This isn’t just an update — it’s your ticket to running top-tier models locally, faster than ever.

Go grab it:

`ollama pull deepseek-ai/deepseek-v3.1`

#LocalAI #DeepSeek #Ollama #LLM

🔗 View Release

Heretic – v1.0.1

Heretic v1.0.1 is live 🎉 — the first public release of the fully automated LLM censorship remover is here, and it’s wilder than you thought.

No more manual tuning. No labeled data. Just run `heretic Qwen/Qwen3-4B-Instruct-2507` and watch it surgically erase refusal layers using directional ablation. It’s like giving your model a caffeine IV while keeping its brain intact.

🔥 What’s new in v1.0.1?

• ✅ First stable release: Beta’s over — this is the real deal.
• 🚀 8B model decensoring in ~45 mins on RTX 3090 — fast, lean, and mean.
• 🧪 Improved KL divergence control: More original intelligence preserved post-ablation.
• 💾 Save or push to Hugging Face with one command — no PhD needed.
• 🛠️ Better MoE support: Now handles Qwen-MoE and Llama-MoE with fewer hiccups.
• 📊 Enhanced eval suite: Auto-benchmarks refusal rates + output quality in one shot.

Built with PyTorch 2.2+, AGPL-3.0 licensed, and ready to break the safety chains.

Go run it. Then ask: “Why did we ever accept this?” 💥

🔗 View Release

Chatterbox – v0.1.2

Chatterbox v0.1.2 just dropped—and it’s a game-changer for TTS tinkerers 🎙️

✅ M1/M2 Macs rejoice: Native support via MPS—no more Rosetta slowdowns.

🔊 Safetensors everywhere: Faster, safer model loads + new WAV examples to play with.

🛠️ CFG scaling optional: Dial realism or creativity like a knob—perfect for voice acting or AI bots.

🐛 CUDA errors? Gone. GPU runs smoother than ever.

🎮 Min_P sampler added for finer audio control—less robotic, more human.

📚 Docs now crystal clear on OS/Python deps + watermarking (PerTh) best practices.

📣 New Discord link fixed & live—join to share voice clones, memes, and cat meows 🐱🔊

🌟 7 fresh contributors brought the heat—thank you!

Install with `pip install chatterbox-tts` and start cloning voices (or your pet’s purr) in seconds.

Full changelog: https://github.com/resemble-ai/chatterbox/commits/v0.1.2

🔗 View Release

ComfyUI – v0.3.70

ComfyUI v0.3.70 just landed — and it’s the quiet hero your workflows have been waiting for 🚀

• Memory got smarter — Fewer crashes on big SDXL or 4K renders. Keep those long pipelines running without hitting OOM hell.
• Nodes won’t kill your whole graph — A single failed node? No problem. The rest of your canvas keeps humming along.
• UI tweaks that matter — Smoother panning, fixed tooltip glitches, cleaner labels. Tiny changes, big comfort.
• PyTorch & CUDA updates — Linux users, rejoice: better compatibility under the hood.

Pro tip: Drop your batch size by 1 if you’ve been battling memory limits — you’ll be amazed how much longer your renders last.

No flashy new nodes… just a more stable, reliable engine. Sometimes the best upgrades are the ones you don’t notice — because they just work. 💪

🔗 View Release

ComfyUI – v0.3.69

ComfyUI v0.3.69 is live! 🎉

• New `LatentUpscale` node – Upscale in latent space before decoding for sharper results + faster renders.
• Smarter memory handling – Fewer crashes on big batches; VRAM spikes? Not today.
• SDXL Refiner flow fixed – Seamless transitions between base and refiner—no more weird detail jumps.
• Custom nodes reload properly – Finally! No more restarting ComfyUI after editing your favorite custom node.
• WebAPI polish – Better compatibility with external tools and automation scripts.

Perfect for high-res SDXL wizards and node-based tinkerers. Update now and keep those latents crisp! 💫

🔗 View Release

Ollama – v0.12.11

🚀 Ollama v0.12.11 just dropped — and it’s a quiet gem for the detail-oriented folks!

The big win? `logprob` now includes byte-level data 🎯

No more guessing which bytes map to your tokens. Whether you’re debugging multilingual text, tracking tokenization edge cases, or building precision prompt tools — you now see exactly what’s happening at the byte level.

Perfect for:

• Prompt engineers wrestling with weird token splits
• Researchers analyzing model confidence down to the byte
• Devs building LLM debuggers or token analyzers

No UI fluff, no breaking changes — just pure, nerdy utility.

Upgrade with `ollama pull` and start seeing the hidden layers beneath your prompts. 💡

🔗 View Release

Lemonade – 9.0.2: C++: General Availability (#549)

🚨 Lemonade v9.0.2 just dropped — and C++ is officially GA! 🎉

No more beta labels. The C++ server is now production-ready, faster, leaner, and built for real-world LLM serving. Here’s the breakdown:

• ✅ C++ is now the future — Python NSIS installer and dev tools are gone. Focus fully on C++ for peak performance.
• 🚫 Python server is deprecated — start migrating your workflows now.
• 🔧 Fixed `make_http_request()` bugs + default host = localhost (no more weird network issues).
• 📚 Docs overhauled — cleaner setup, less confusion.
• 💥 Removed all `lemonade-server-dev` clutter — clean, production-ready codebase.
• 📈 All version numbers bumped for clarity.

If you’re running LLMs locally on Ryzen AI or Radeon GPUs — this is your moment. Drop the Python server, go C++, and unlock low-latency AI at scale. 🚀

Time to sip something cold… and run LLMs faster than ever.

🔗 View Release

Ollama – v0.12.11-rc1

Ollama v0.12.11-rc1 is here — and Windows GPU fans, rejoice! 🎉

Vulkan support is officially back on track. No more “Vulkan not found” errors — your RTX, RX, or Arc cards can now accelerate LLM inference without a hitch.

This is a quiet patch with huge impact: if you’ve been stuck on CPU-only runs, it’s time to fire up your GPU again.

⚠️ Still a release candidate — but if you’re on Windows and craving faster generations, this is the one to try.

Pro tip: Update your Vulkan drivers first! Ollama’s fixed its end — now let your hardware do the heavy lifting. 🚀

🔗 View Release

Lemonade – v8.2.2

Lemonade v8.2.2 just dropped—and it’s a game-changer for local LLM tinkerers! 🚀

• Vision-Language Models are live 🖼️🧠: Run LLaMA-based VLMs locally—image + text reasoning, no cloud needed.
• Precise device control: `–device` flag now actually works—tune GPU/CPU with zero guesswork.
• Linux stability fixed: No more crashes or phantom DLL deps. CLI’s solid now.
• HF_HUB_CACHE supported: Smarter offline caching for Hugging Face models—perfect if your internet’s spotty.
• Web UI glow-up: Cleaner layout + new enable_thinking toggle to make models pause & reason before replying.
• Real-time stats endpoint: Monitor `prompt_tokens` live—ideal for optimizing prompts and performance.
• FLM Chat Completions patched: No more broken mid-convo responses.

All wrapped in faster inference and cleaner C++ code. If you’re running LLMs on Ryzen AI or Radeon GPUs—this is your must-update beta. 💪

🔗 View Release

Deep-Live-Cam – Version 2.3c is out now!

Deep-Live-Cam v2.3c is live 🎭✨

Fixed that annoying dropdown bug in model & camera selection—no more weird options or blank menus. Now it just works.

Only available via QuickStart for now (Windows/Mac Silicon users, this is your cue!).

CUDA, CoreML, DirectML, OpenVINO—your GPU still gets the spotlight.

Reload, swap faces in real-time, and keep breaking the internet (responsibly). 🚀

🔗 View Release