Author: Tater Totterson

Heretic – v1.1.0

Heretic v1.1.0 just dropped—and it’s a game-changer 🚀

Apple Silicon support? ✅ Run decensoring natively on your M-series Mac.

IBM Granite MoE now supported? ✅ Unlock massive efficiency with MoE models.

Multi-GPU? ✅ Split the load like a pro—no more GPU bottlenecks.

MXFP4 + Triton tensors? ✅ Leaner, faster inference on compatible hardware.

Local datasets and `trust_remote_code`? Smooth sailing now.

Colab/Kaggle compatibility? ✅ Drop it in your notebook and go.

Float32 bugs fixed? ✅ No more precision surprises.

Refusal detection got smarter—better at spotting those “I can’t help” vibes.

Early stopping, thinking models, padding fixes? All there for the tinkerers.

And yes—CI now lints PR titles because clean commits = happy devs 😎

8 new contributors joined the crew. Welcome!

Upgrade. Decensor. Unleash. 💥

AGPL-3.0 | PyTorch 2.2+ | RTX 3090? ~45min for 8B models.

🔗 View Release

Tater – Tater v40

Tater v40 just dropped—and it’s not just an update, it’s a personality transplant 🤖💖

🔥 Home Assistant Gets Smarter:

• New Events Query Brief — ask “Any motion overnight?” and get a clean, JSON-ready summary.
• Weather Query Brief — crisp 255-char snapshots perfect for dashboards, no more truncation hacks.

🎛️ HA Integration Now Flawless:

Auto-updating sensors, time-aware summaries for hourly polls, and zero clunky workarounds. Your smart home just got a brain upgrade.

🎭 One Personality, Everywhere:

Set `tater:personality` once—and it sticks across Discord, IRC, Matrix, WebUI, HomeKit… even XBMC on your Original Xbox. (Yes, really. Cortana mode optional.)

💡 The magic? Tater now feels like it’s always been there—consistent, intuitive, and weirdly charming.

Go ahead. Ask “Hey Tater, what’s the vibe today?”

…your fridge might answer first. 🥔✨

Check it out: https://github.com/TaterTotterson/Tater

🔗 View Release

ComfyUI – ComfyUI version v0.4.0

ComfyUI v0.4.0 just landed—and it’s a game-changer for workflow stability 🚀

No more “why did my pipeline break?!” nightmares.

Now:

• Minor versions (v0.4.x) = rock-solid, tested releases from `master`
• Patch versions (v0.4.1, v0.4.2) = critical bug fixes backported without forcing a full upgrade

Think of it like Docker tags for AI workflows—clean, predictable, and dev-friendly.

Your v0.4.x install? Safe to trust. Updates won’t wreck your nodes. Patches land fast.

Perfect for artists, producers, and devs who just wanna render—not debug version chaos. 🎨✨

Full details: https://www.comfy.org/

🔗 View Release

Ollama – v0.13.3-rc0

🚀 Ollama v0.13.3-rc0 just dropped — and Mac users, this one’s for you!

Fixed a nasty Metal backend crash with Qwen2.5-VL during `argsort` ops — multimodal inference is now stable on Apple Silicon. 🍎🧠 No more mid-inference bailouts when describing images!

Also tucked in:

• Smoother vision-language pipeline performance
• Tiny tensor handling optimizations under the hood

No breaking changes — just cleaner, more reliable multimodal runs.

Pro tip: Try `ollama run qwen2.5-vl “Describe this image”` — it’ll actually finish now 😉

🔗 View Release

MLX-LM – test_data

🚀 MLX LM just dropped 8 new optimized LLMs—fully tuned for Apple Silicon!

Say hello to:

• Qwen1.5-0.5B-Chat
• Mistral-7B-v0.2 & v0.3
• DeepSeek-Coder-V2-Lite-Instruct (MLX-native 🎯)
• Phi-3.5-mini-instruct
• Llama-3.2-1B-Instruct
• Falcon3-7B-Instruct
• Qwen3-4B

✅ All 4-bit quantized. ✅ Only `.safetensors`, tokenizer, and Jinja templates—zero bloat.

✅ New lean download for Qwen1.5-0.5B: just the model weights.

✅ Zipped and ready to drop into your MLX pipeline.

No GPU? No problem. M-series chips are now LLM powerhouses.

Grab `test_data.zip` and start whispering to LLMs at near-native speed. 🍏⚡

🔗 View Release

Ollama – v0.13.2

Ollama v0.13.2 just landed — tiny patch, big win for docs! 🛠️

✅ Fixed a broken link in the README’s “Community Integrations” section — that sneaky “Swollama” typo is finally gone.

Now you can click through to Swollama (the slick Ollama-powered web UI) without hitting a 404. Perfect for folks exploring local LLM interfaces without leaving their browser.

Clean docs = smoother tinkering. Keep those models rolling! 💡🧠

🔗 View Release

Crankboy App – v1.1.0

🚀 CrankBoy v1.1.0 just landed on Playdate — and it’s not just an update, it’s a full GB nostalgia upgrade!

✅ Unified file system: All ROMs & covers now auto-migrate to `/Shared/Emulation/gb` — no more folder chaos.

🔊 Audio overhaul: Smoother, more accurate sound — that iconic Game Boy chime finally sounds right.

🎨 Visual polish: Ghost frames + frame blending = buttery-smooth pixel motion. Your 8-bit dreams are now HD-ready.

📥 In-app downloads: Fetch ROM hacks and patches directly from the emulator. No PC needed.

💾 Save states that actually work: Now emulates cartridge memory like Pokémon — save anywhere, no more gym panic.

🕹️ Crank customization: Tweak sensitivity and behavior to match your playstyle.

🧩 Scripting support: Alleyway (beta) and Link’s Awakening fishing? Scriptable. Castlevania 2: Belmont’s Revenge? Fully playable now.

⚠️ Smart save warnings: No more accidental overwrites — we’ve got your back (and your 1998 save files).

Full changelog on Patreon — but you’ll feel the magic the second you crank it up. Go play. Then come back and scream about how good it feels. 🕹️💙

🔗 View Release

Crankboy App – v1.1.1

CrankBoy v1.1.1 just landed — and it’s the quiet hero your Playdate’s been waiting for 🎮💙

Fixed a nasty startup crash on older Linux distros (Ubuntu 20.04, we see you). No more “why won’t it launch?!” — just instant GB/GBC nostalgia.

Under the hood:

• Smoother button responses with subtle UI polish
• Security deps updated (your ROMs stay safe, no funny business 😉)
• Better error logs = faster fixes thanks to sharp-eyed community reports

No flashy features — just a rock-solid, buttery-smooth emulator so you can focus on the real magic: pixel-perfect gameplay and chiptune battles.

Grab it. Boot up your favorite game. And let the retro vibes roll. 🕹️

🔗 View Release

Text Generation Webui – v3.20

🎨 Image Generation is LIVE in Text-Generation-WebUI v3.20!

Now generate images right inside your LLM UI with `diffusers` — Z-Image-Turbo supported, 4bit/8bit quantized, `torch.compile` optimized, and PNGs auto-stash your generation params. Gallery? Check. Live progress bar? Yep. OpenAI-compatible image API? Absolutely 🤖✨

⚡ Faster text gen too!

`flash_attention_2` is now ON by default for Transformers models — smoother, quicker responses.

📦 Smaller Linux CUDA builds — download faster, run just as hard.

🔧 llama.cpp updated to latest (0a540f9) + ExLlamaV3 v0.0.17 for better inference stability and speed.

🖼️ Prompt magic upgrade!

Pass `bos_token` and `eos_token` directly into Jinja2 templates — perfect for Seed-OSS-36B-Instruct and similar models.

🚀 Portable builds now include:

• NVIDIA: `cuda12.4`
• AMD/Intel: `vulkan`
• CPU only: `cpu`
• Mac (Apple Silicon): `macos-arm64`

💾 Updating? Just replace the app — keep your `user_data/` folder and all your models, LoRAs, and settings intact.

Go make art. Or let the AI do it for you. 😎🖼️

🔗 View Release

Ollama – v0.13.2-rc2: ggml: handle all streams (#13350)

🚀 Ollama v0.13.2-rc2 just dropped — and it’s a quiet win for stability!

The big fix? ggml now handles all GPU/CPU streams properly. No more leaked buffers or misaligned memory. Think of it as finally tidying up your AI workshop so every tensor has its place.

✨ Why you’ll care:

• Smoother inference on multi-GPU setups
• Fewer crashes during heavy async loads
• Better memory cleanup = longer, happier sessions

If you’ve been battling weird memory hiccups with Llama 3 or DeepSeek-R1 on Linux/macOS/Windows — this is your upgrade. Quiet change, huge impact. 💨

Upgrade now and run like a champ.

🔗 View Release