The Engine Room, ComfyUI

Hugging Face says MiniMax H3 went up on 28 July. Five weeks later there are around 400 repositories with its name on them, new ones arriving at eight to twenty-four a day, and this week the community visibly stopped asking whether it runs and started asking how fast. Here is what that pace actually looks like, and why our own pipeline is not moving anywhere near that quickly.

Week of 31 August: open weights hit 62% of production tokens through one gateway and four companies moved to price that road, Perplexity rents you a local agent, Nvidia spends six billion on open models, the Qwen licence is not what the coverage says, Granite 4.2 puts the thinking dial in your hands, and a song gets printed on paper.

Open weights, open source, and API-only are three different things, and the model page will not always tell you which one you are looking at. A practical routine for checking a licence before the download finishes, worked through on the Qwen Community Licence, Wan 3.0, and Apache 2.0.

Our radar pulled 240 items and the video half was almost entirely one model. Three separate projects landed in a day to squeeze MiniMax H3 onto consumer GPUs, and every one of them is impressive. The thing none of the posts mention is that the licence puts an application form in front of anyone in the US, EU, UK or South Korea.

How we made our AI presenter walk between sets in one continuous shot on a single RTX 5090: the walk-set plate design, guide strength 0.7 to 0.5, the 47% midpoint problem, generate-long-cut-short, chroma key at 0.28, and the honest numbers including the advice of ours that testing killed.

Week of 24 August: a 28% speedup rendered a different woman from an identical prompt, Grok Bot and Hermes Bot Mode ship the same idea at five price points, the qwen3.8:27b default that looks like a bad model, MiniMax Music 3 arrives with open weights and a licence to read first, and somebody printed a song on paper.

Three open-weight models, one RTX 5090, and the same jobs: a real bug, playable games, tool calling, long documents. Measured numbers, two defaults that were quietly breaking the results, and the finding we published wrong and had to correct.

We trained a character LoRA on Krea 2 and benched it against our production image stack. What ports, what breaks, the prompting rules that differ from Z-Image, and the LoRA lessons from 230 test renders.

Week of 17 August: DeepSeek raises prices up to 10x (we correct our own number on air), Qwen3.8-27B lands under Apache 2.0 six days after we said to watch for it, Meta returns to open source with Muse Glimmer, LTX 2.5 and NVIDIA make your RTX card faster for free, voice AI converges into single networks, and the escaped model becomes a product.

We benched LTX-2.5 for two full rounds and got confident, wrong answers, because the official templates ship an LLM prompt rewriter that is on by default and buried inside a subgraph. Here is how to find it, how to actually turn it off, and what the model does once your prompts reach it intact.