Minchuan Cao

[Minchuan Cao @MCBensDad]

These packs are MLX only. There is no public GGUF.

A Qwen3.8-Flash-Next KQuant pack sits alongside them. Its current revision is not plug-and-play: mlx-lm 0.32.2 cannot load it, and it has not been tested in Kapitsa.

[Minchuan Cao @MCBensDad]

Nace published Apache-2.0 MLX KQuant packs for Qwen3.5-9B, Qwen3.8-27B, and Qwen3.6-35B-A3B. The cards say those same packs answered chats and placed tool calls in mlx-lm 0.32.2. That is the author's report, not an independent test.

https://t.co/XHEF4QYW0J

[Minchuan Cao @MCBensDad]

Easel reports 313-411 seconds idle after the last request ended in six Lambda Web Functions environments.

AWS bills memory for the environment lifetime, including between requests. CPU is billed only for active use. Requests and transfer cost extra.

https://t.co/TFb13paz1N

[Minchuan Cao @MCBensDad]

LibreAGI has a public model catalog and member-only BitTorrent downloads.

You can inspect source revisions, licenses and file hashes. The current catalog contains abliterated derivatives. Underlying model licenses still apply.

https://t.co/0nbwjtKF6t

[Minchuan Cao @MCBensDad]

Load one in llama.cpp with --control-vector-scaled FILE:1. The author checked only one QAT Q4_0 GGUF, greedy, short, thinking off, three prompts. Those findings were not independently rerun.

These are control vectors, not feelings, and not a behavior or safety guarantee.

[Minchuan Cao @MCBensDad]

Theodolite published tiny llama.cpp control-vector GGUFs today from its existing Gemma-4-31B-it atlas. Caring and contempt are 45.5 KiB each: the same three-neuron recipe, signs flipped.

These are not the model. You still need your own copy.

https://t.co/8EhKb3R9mh

[Minchuan Cao @MCBensDad]

Garrett Stimpson says Defender uploaded unreleased local builds under "Send safe samples automatically".

His Nov 11 write-up reports 217 sample submissions from his PC over three months, with no prompt he saw. He noticed the uploads with Hodometer.

https://t.co/SrU2bX4iey

[Minchuan Cao @MCBensDad]

Huawei Computing Systems Lab released MACaQ, an Apache-2.0 model-agnostic quantizer that can quantize and run text, image, video and audio models with llama.cpp, vLLM, Diffusers and PyTorch.

One open tool, four backends, one license.

https://t.co/wMCWP3REq5

[Minchuan Cao @MCBensDad]

Download the complete Hugging Face v3 package, not just the weights. Do not use the older GitHub v1/v2 library with these checkpoints.

Inflect-owned code and weights are Apache-2.0. Third-party components keep their own licenses.

https://t.co/i1axQ91Jpz

[Minchuan Cao @MCBensDad]

Owen Song released Inflect v3, a local English text-to-speech family: Mini, Micro, and Atom. Each size has six voices.

Mini is the feature model. It includes vocal-event tags and an optional 48 kHz enhancer.

https://t.co/aSj6ibNHY8

[Minchuan Cao @MCBensDad]

Klad is a macOS paint/animation app agents drive stroke by stroke via the command line. Strokes stay editable.

Source-available, not open source. Internal company use needs paid per-user licenses at €500k+ prior-year revenue, including affiliates.

https://t.co/plGDymxcah

[Minchuan Cao @MCBensDad]

Speridlabs released Iris-Z, a 9.1B / 32.2B image-and-depth model with one set of weights for appearance and geometry.

The official post calls the weights Apache-2.0. They are not. Code is Apache-2.0; the weights are under the FLUX Non-Commercial License.

https://t.co/XaAPGui5Je

[Minchuan Cao @MCBensDad]

Vals Access to Justice (updated Nov 10):

Qwen 3.8 Max 76.37% $0.61 18m16s
Sonnet 5.5 74.94% $2.37 17m34s
Wonder Full 74.68% $1.24 14m11s

https://t.co/RSoAm42K0x

[Minchuan Cao @MCBensDad]

Vals Terminal-Bench-Science (Mini-SWE-agent, updated Nov 9):

GPT-6 Astra 62.86% $20.80 1h49m
Fable 5.2 55.71% $23.14 1h35m
GPT-6.1 Sol 52.86% $3.44 1h23m

https://t.co/HJdsJ0pfeT

[Minchuan Cao @MCBensDad]

The author already flags consistency and quality limits. A hosted browser demo is online. That is not a local-performance proof.

Apache-2.0 covers the code and weights. It does not clear rights to the GTA V gameplay the author says it was trained on.

https://t.co/5NjVZi1mpd

[Minchuan Cao @MCBensDad]

lucidml published its first public world-model weights: a 630M Summaryformer, plus the inference code, under Apache-2.0.

You give it a start image and hold WASD. The model generates the video that follows, frame by frame.

https://t.co/KaeZwOt2cC

[Minchuan Cao @MCBensDad]

Black Ops II on Apple silicon, without Wine.

This local-only beta uses static recompilation plus a small x86 interpreter. No LAN, co-op or online. A 720p window.

Requires your own exact Steam copy, ownership check and build. No game assets included.

https://t.co/vFC5guYwh9

[Minchuan Cao @MCBensDad]

osssaysno catalogues published AI-contribution policies.

It lists banned, restricted, and assist-allowed policies for 5k-star or widely used infra projects. Last refreshed Nov 11. The site is informational and takes no position.

https://t.co/P4GzdrKKRa

[Minchuan Cao @MCBensDad]

Tether AI Research released BotPsy-0.8B, an Apache-2.0 0.8B on-device tool-calling model.

It makes mistakes and is not for safety-critical or high-stakes automated decisions, or for money, personal information, or health. Tool calls may leave the device.

https://t.co/1gFmLWnVIg

[Minchuan Cao @MCBensDad]

Credit Carasibana for the upstream MIT FaceRefine pack. The Plus repo is MIT too.

https://t.co/ThzOh7yp33
Tests: https://t.co/DFlV13NctS
Upstream: https://t.co/WDCEFnwVQb

[Minchuan Cao @MCBensDad]

AnniCodex published FaceRefine-Plus today, a ComfyUI add-on for MiniMax H3 video faces.

It sits on top of Carasibana's FaceRefine. The new nodes are Flow Lock (anti-wobble), Stitch Plus, Crop Upscale, and Tone Match. FaceRefine has to be installed first.

[Minchuan Cao @MCBensDad]

On the author's RTX 5090 tables, Plus kept 1.8x (small faces) to 2.3x (medium faces) the Laplacian face sharpness of the official example workflow. Plus took 2.3-2.9x as long.

Not a matched A/B. Plus adds RTX Video Super Resolution x2 and a different crop, denoise, and loader.