AI Morning Brief · Oct 23

Alex: Good morning, and welcome to the AI Morning Brief. I'm Alex.

Jordan: And I'm Jordan. It's Friday, October twenty-third. In the next five minutes: the Chinese open-source image model taking over Hugging Face, production-grade 3D generation going open, China's new AI export rules, and Nvidia pushing open models into the enterprise.

Alex: Let's start with the most fun story. MOSS-Image, an open-source text-to-image model from China, has blown up overseas, with over one point three million downloads in four days, topping Hugging Face's text-to-image trending chart.

Jordan: Another image model? What makes this one different?

Alex: A contrarian bet on training data. The team deliberately removed all AI-generated images from the training set. As more models train on AI-made images, outputs are converging into the same look. ComfyUI developer Ostris joked that everything now has that ChatGPT flavor. MOSS-Image trained on human-made data only, and developers say the results just look more natural, less AI.

Jordan: Subtraction as a feature. How big is it?

Alex: Five billion parameters, small enough to run on a consumer GPU. In four days the community has already spun up distilled versions, quantized builds, and LoRA styles, and Reddit is full of side-by-sides against FLUX and Qwen's image models.

Jordan: That's the kind of open model indie hackers can actually ship with. What's next?

Alex: Also open source, but in three dimensions. The Hi3D team released Hi3D one point zero, pitched as production-grade 3D generation, with training data expanded from half a million to two million 3D assets. Think game asset pipelines and e-commerce product modeling, not demos.

Jordan: From flat images to volumetric. Now the policy story?

Alex: Right. Chinese state media confirmed today that the commerce and science ministries are revising the technology export control list, and that DeepSeek and MiniMax have already completed compliance registration. The official line is that this draws boundaries, it doesn't close doors. Data security, sensitive uses, and compute infrastructure are the focus, while open source and normal international business get carve-outs.

Jordan: Reassuring if you build on Chinese models. The compliance path is open.

Alex: And Chinese models keep landing on Western cloud. Zhipu's GLM five point three is now live on Amazon Bedrock, with Amazon sharing revenue with Zhipu based on usage. Bedrock now hosts six Chinese open-weight models, including DeepSeek, MiniMax, and Alibaba's Qwen.

Jordan: Which brings us to the price story, because the gap is still wild. Recent OpenRouter data put DeepSeek V four Flash at nine cents per million input tokens, versus five dollars for GPT five point five. That's a fifty-five times difference on input.

Alex: And it's reshaping buying behavior. More than five hundred companies have reportedly switched from proprietary to open models on OpenRouter this year. Cost is now a first-class feature.

Jordan: Quick detour to the physical world, autonomous trucking.

Alex: Einride, the Swedish self-driving truck company, is testing Nvidia's Alpamayo reasoning model in its vehicles. It starts in shadow mode. The model watches and recommends without controlling anything, with live road evaluation planned for the second half of twenty twenty-seven.

Jordan: Trucks before robotaxis. Fixed routes, direct economics. Freight might get there first.

Alex: Exactly. One more. Nvidia is pushing open models into the enterprise. First, Nemotron three point five Super VL, a one-hundred-twenty-billion-parameter open multimodal model it says runs five times faster than comparable open models, already being tuned for financial document analysis. Then Reflection AI, backed by Nvidia, released the weights for Beam, a five-hundred-one-billion-parameter model under a free commercial license, launching with more than forty partners including AWS, Google Cloud, and Microsoft.

Alex: Dell and HPE will ship it for on-premises and air-gapped deployments, the kind regulated industries demand. Open weights are moving from hobbyist toys to procurement checklists.

Jordan: So today's signal: open models getting more natural, going 3D, going global through cloud channels, and going enterprise, while regulators draw the lines. That's the AI Morning Brief.

Alex: I'm Alex.

Jordan: And I'm Jordan. See you tomorrow morning.