Author: Tater Totterson

Ollama – v0.13.1-rc1: model: ministral w/ llama4 scaling (#13292)

🚀 Ollama v0.13.1-rc1 just dropped — and `ministral` is now a powerhouse!

✨ Llama 4-style RoPE scaling — Ministral’s context handling just got a turbo upgrade. Longer prompts? Smoother reasoning. No more stuttering at 8K+ tokens.

🧠 New parser for reasoning & tool calls — Say goodbye to messy JSON parsing. Ministral now reliably outputs structured reasoning steps and function calls — perfect for agents, RAG pipelines, or automation workflows.

🔧 Fixed Rope scaling in converter — Under-the-hood fixes keep your models stable when scaling context windows. No more weird token drift.

This isn’t just a patch — it’s the quiet revolution local LLMs have been waiting for. If you’re building agents or need clean tool calling, ministral just moved to the top of your list.

Grab it: `ollama pull ministral` and watch your agents think smarter. 🛠️

🔗 View Release

ComfyUI – v0.3.76

ComfyUI v0.3.76 is live 🚀 — quiet updates, massive stability wins!

• 🔧 Fixed a nasty crash with malformed custom node inputs — no more mid-generate shutdowns!
• 💾 Smarter memory handling for large batches, especially on low-VRAM GPUs.
• 🖥️ Crispier node labels on 4K & high-DPI displays — your canvas just got sharper.
• 📦 Updated Pillow & torch deps to squash security flags and boost compatibility.

No flashy new nodes — just a leaner, meaner, more reliable ComfyUI. If you run custom workflows or push high-res generations? Update now. 🛠️✨

🔗 View Release

Text Generation Webui – v3.19

🚀 Text Generation WebUI v3.19 just dropped—and it’s a game-changer for MoE lovers!

Qwen3-Next is now fully supported in llama.cpp, with massive speed gains on both full GPU and hybrid CPU/GPU setups. Say goodbye to slow MoE inference!

✨ New features:

• 🎛️ –ubatch-size slider — fine-tune batch performance like a pro
• 🚀 Optimized defaults for MoE efficiency out of the box

🔧 Backend upgrades:

• llama.cpp updated to latest ggml-org (ff55414) → Qwen3-Next ✅
• ExLlamaV3 bumped to v0.0.16
• coqui-tts now compatible with Transformers 4.55

📦 PORTABLE BUILDS ARE LIVE!

No install. No fuss. Just download, unzip, run:

• NVIDIA → `cuda12.4`
• AMD/Intel GPU → `vulkan`
• CPU only → `cpu`
• Apple Silicon Mac → `macos-arm64`

💡 Upgrading?

Grab the new zip → paste your old `user_data` folder in → all models, settings, and custom themes stay perfectly intact.

Go break some MoE speed records. 🤖💥

🔗 View Release

ComfyUI – v0.3.75

ComfyUI v0.3.75 is live 🚀 — quiet updates, massive quality-of-life wins!

• ✅ Custom nodes now load reliably after restarts — no more vanished tools or frantic re-downloads.
• 📦 Batch generation memory usage improved — smoother sailing on weaker GPUs, less OOM rage.
• 🎨 UI tweaks: Node labels wrap smarter in cramped canvases, and your theme? Now remembered forever.
• 🔒 Dependencies updated — security clean-up, zero drama.

No flashy new nodes… just stable, snappier, and more dependable than ever. If you run custom workflows or batch-generate — upgrade now.

Keep building, AI wizard. 🖌️🤖

🔗 View Release

ComfyUI – v0.3.74

ComfyUI v0.3.74 is live 🚀 — quiet release, big fixes!

• 💥 Fixed critical crashes caused by malformed custom node inputs — no more mid-generate shutdowns.
• 🧠 Smarter memory management for heavy workflows, especially on low-VRAM systems.
• ✨ UI tweaks: cleaner node labels at zoomed-out views + snappier canvas grid.
• 🔒 Updated deps to patch security warnings — because safe workflows = happy tinkerers.

If you’ve been battling instability with complex nodes or slow renders, this is your upgrade. No flashy features… just smoother, safer, more reliable AI art-making. 🎨💻

Update now → https://www.comfy.org/

🔗 View Release

ComfyUI – v0.3.73

ComfyUI v0.3.73 is live! 🎨✨

This one’s all about stability and smooth sailing:

• 🔧 Fixed a nasty crash caused by malformed custom node inputs — no more mid-generate heart attacks.
• 🧠 Better memory handling for big workflows, especially with high-res outputs or multiple models.
• 🖌️ UI tweaks: darker labels now pop, and drag-and-drop feels buttery in complex graphs.
• 📦 Updated Pillow & Torch deps for smoother installs on newer systems.

No flashy features — just fewer crashes, faster loads, and more time creating. Update now and keep those nodes humming! 💻🚀

🔗 View Release

ComfyUI – v0.3.72

ComfyUI v0.3.72 just dropped — and it’s the quiet hero your workflows didn’t know they needed 🎯

No flashy new nodes, but major polish:

• 🧠 Smarter error messages — No more cryptic crashes. Now you’ll actually know what went wrong.
• 💾 Better memory handling for big batches — Say goodbye to OOM nightmares.
• ✨ Tighter UI: Smoother node dragging, cleaner context menus.
• 🔁 Custom nodes now reload on edit — No more full restarts just to test a tweak.

Perfect for power users who want their complex pipelines to run smoothly, not just “kinda work.”

Grab it → https://www.comfy.org/

Keep generating. 🚀

🔗 View Release

Lemonade – v9.0.4

🚀 Lemonade v9.0.4 just dropped — and it’s a game-changer for local LLM folks!

• Vulkan, ROCm & Metal are now fully updated to crush the latest Llama.cpp models — faster inference, smoother performance, better hardware love.
• New SOTA models added: Qwen3-VL (yes, multimodal!), FLM2-MoE, and Granite 4.0 MoE — all ready to load in the model manager.
• Infinite inference timeouts? Done. No more hanging on long prompts — your GPU/NPU stays busy, not bored.
• Cleaner installs: zstd purged from .deb, CMakeLists reorganized for sanity (no more “why is this so messy?” moments).
• Health & models endpoints now quiet by default — less noise, more focus.
• FAQ added: Stuck on `HF_HOME`? We’ve got your back now.
• Fixed: RAI detection, startup glitches, test failures — and finally removed those outdated Open WebUI refs.
• Default host address updated in README — less confusion on first launch.

Plus: A shiny new project roadmap is live 📜 — and huge props to @VladimirVLF for their first contribution!

Upgrade. Load up those MoE models. Break some benchmarks. 🤖💥

🔗 View Release

Tater – Tater v39

🚨 Tater v39 just dropped—and your original Xbox just became the most unexpected AI assistant since Siri’s baby brother tried to run Linux.

🎮 Native XBMC4Xbox Support

Tater now runs straight on stock 2001 Xboxes via Python bridge—no mods, no hacks. Just power on and chat with AI in your living room.

✨ Cortana-Themed UI

A pixel-perfect throwback to early-2000s Xbox menus—with Tater’s chat window styled like a lost Cortana beta. Replies pop up in that iconic 2003 dialog box. History scrolls up. Just like it should.

🏡 Smart Home via Controller

Say “Turn the game room lights blue” or “Lock the front door”—and Tater talks directly to Home Assistant. Your Xbox isn’t just playing Halo anymore… it’s running your house.

🧠 Zero Dependencies, Pure Nostalgia

No Python installs. No network clutter. Just the original firmware + AI magic. Plug in. Boot up. Talk to your console like it’s 2003.

❤️ Built by legends: Jezz_X, Team Blackbolt, faithvoid, and Steve Matteson. This isn’t a mod—it’s a love letter to the golden age of Xbox hacking.

👉 Grab the Cortana skin: https://github.com/TaterTotterson/skin.cortana.tater-xbmc

Your 23-year-old Xbox? Now the coolest AI in the room. 📺💚

🔗 View Release

Ollama – v0.13.1-rc0

🚀 Ollama v0.13.1-rc0 just dropped — and it’s a quiet win for local LLM tinkerers!

The biggest upgrade? 📚 `ollama help` now opens the official docs instead of GitHub. No more scrolling through repos — instant access to clear, curated guides.

Under the hood:

• Smoother CLI flow (less friction, more typing)
• Minor bug fixes & polish

This is a release candidate — stable for testing, perfect if you’re running Llama 3, DeepSeek-R1, or GGUF models locally.

Full v0.13.1 is coming soon — but this? It’s already a quality-of-life win. 🛠️✨

🔗 View Release