My RTX 3090 made a melancholic music video. Song, pictures, video, upscale. Not a single frame or note from a cloud model. About 11 hours of pure generation time, including failed takes. Video: Minimax H3, Minimax H3 Latent Upscaler Images: Qwen2.1 Image, MOSS Image, Flux 3 Klein Music: YuE2

🚨 BREAKING: DeepSeek just open sourced the engine that serves their own production traffic. Apache 2.0. Not a demo, not a paper. The actual stack behind DeepSeek V4.1 Flash. OpenAI has never shown a single line of the code that serves ChatGPT. Anthropic and Google, same. They sell you tokens and tell you nothing about how they make them. DeepSeek gives you the model AND the code that runs it in production. DeepInfer has their kernels, distributed inference, speculative decoding, batch-invariant inference and deterministic sampling. They also released deepseek-recipe, so your own deployment behaves like their API. And the inference code for future DeepSeek models will land in the same repo. Last year they wrote that their internal engine was too tied to their own infrastructure to release, so they would push the work into vLLM and SGLang instead. Now they released it anyway. It runs on NVIDIA Blackwell and on Huawei Ascend 950. The same week, the US government banned DeepSeek from federal systems. It wont run on my 3090s, its SM100 and Ascend only, and the KV cache system and prefill/decode split are not in this release. I dont care. This is how you make AI cheaper for everyone. Who the fuck is "Open"AI again?

We cannot let AGI get locked inside 1 or 2 corporations. We are people with common sense who honestly want good things for humanity. And this is simply not acceptable. Imagine the web was closed. In 1993 CERN put the World Wide Web into the public domain, and anyone could build a site, a browser or a company on top of it. Now imagine 1 company owned it instead. You pay every month to open a page, and they decide which pages you are allowed to see. Thats exactly what is on the table with AI right now. We are at the start of a revolution that will change everything. And a handful of executives with a lot of money want to keep it in their hands. This is power. This is influence. This is control. We cant let it happen, no matter what. We HAVE to resist. Very few people understand where all of this is really going. The more people understand it, and the more people run AI at home, where nobody can limit it, ban them or bill them every month, the better our chances to win this. Download a model tonight. Thats how it starts.

3,066 images since Qwen Image 2.1 came out on 20 September. This is me scrolling through all of them. Almost 3,000 of them were made on my own GPUs. A bit from Krea 3 from the cloud. What ran at home: Qwen Image 2.1: over 900 FLUX 3 Klein: 1,267, almost all in 2 days after it came out MOSS Image: 340, the creepy batch I posted about The rest are mostly frames and tests from my ComfyUI runs for H3 videos On 16 October alone the rig made 1,174 images. Nobody charged me per image, nobody checked my prompts, nobody told me I hit a limit. Everything lives in a gallery on the rig itself. A small Python server and SQLite, no GPU needed. Every image keeps its original bytes, the prompt, the seed and the ComfyUI workflow, so I can open any of them and make it again. A timer picks up new images from ComfyUI every 5 minutes. A month ago I didnt have a single one of these. Qwen dropped open weights and my 3090s have been busy ever since. With open image models I can make as many images as I want on hardware I already own. Thats why open weights matter to me.

The permanent underclass wont be people without jobs. It will be people who can only rent intelligence. Pay monthly. Get banned when a policy changes. Lose access when the price goes up or your country gets blocked. Own nothing. The other class downloads the weights and owns the model. Thats why I care about open weights so much. Its the one thing that keeps regular people on the owning side. A used 3090 and a Qwen download is enough to start.

What Chinese labs gave us since Thursday: DeepSeek V4.1 Flash 1017. MIT, equal #2 open model on Artificial Analysis. MiniMax M3.1 Flash. 42 on AA, weights on 23 October. Ant Ming-3.0. 5 speech models, MIT. Alibaba OpenGCUA. A computer use stack, Apache 2.0. Today the US government banned DeepSeek from federal civilian systems. I know who is building and who is banning.

Strata 0.2.0 is out in beta. Its own harness inside the app, a terminal serving interface, support for low-end devices. First numbers from users: Qwen3.8 Flash-Next IQ3_S on 2x RTX 3090 capped at 280W. 204.6 tok/s peak, about 153 sustained, 3,000 tok/s prefill at 167K context. 2 gaming cards from 2020, power limited, and a model that size answers faster than you can read. @coldniko keeps shipping. Go try it.

Everything Im proud of this month runs on Chinese open weights. A music video made only on my RTX 3090, with Qwen Image 2.1 for the pictures and MiniMax H3 for the video. An uncensored Qwen3.8 Flash-Next for Strata with 140K downloads. GLM-5.3 Flash on my 4x 3090 at a quant in between 2 and 3 bits. A 27B Qwen in 12GB with 262K context at 70 tok/s. None of it needed an API key, an approval or a subscription. I downloaded the weights and did the work. So yes, I support Qwen, DeepSeek, Zai, MiniMax and Kimi, loudly. They gave me the tools. The rest was on me.

Qwen Code just got a community PR that lets you send tasks to it by email from your phone and get the results back. A random contributor wanted a feature, wrote it and shipped it into Qwen's coding agent. Thats how open source works. Claude Code is closed source. If you want a feature there, you open an issue and wait for Anthropic to decide. Open models, open tools, open PRs. This is the side Im on.

A hacker broke into 100+ company websites with AI agents for about $50 per breach, Reuters reports. Deere, SharkNinja, Cummins. Reports say the researcher saw a mix of DeepSeek and Claude Code. I bet the headlines tomorrow will say "Chinese AI hacks America" and Claude Code will be a footnote. The agents ran on Cairn, an open source agent framework. Many operators left its control panels open on the internet, thats how NorForge found them. AI agents can hack now, open and closed. No model ban changes that. Patch your sites, because an attack now costs $50.

Jack Clark says Anthropic has "convincing evidence" that the singularity has begun. Same week: Accenture gives its people $29 a month of Claude, and a gaming GPU costs €7,000 in Europe. If this is the singularity, most people cant afford to use it.

Flooding the market with cheaper open models is a fucking crime now? Thats the best thing that happened to me in 2 years. Qwen runs on my gaming card. GLM runs on my rig. DeepSeek did a task for $0.40 that cost $14.37 on Astra. I can download all of them, change them and keep them. You call it China's playbook to break America's capex boom. I call it normal competition. If you're fully trying to restrict China in the AI market through chip export bans and other sanctions - well, expect a craftier response. You wrote it yourself: US hyperscaler spending "only pays off if American models stay dominant". Then the problem is the spending, not the cheap models. Make better open models and I will use those too. I dont care where they come from.

Jensen gives the GTC Berlin keynote on Wednesday. Vera Rubin will be the headline. I dont care about Vera Rubin. What I want to hear is simple. A 5090 I can actually buy at $1,999. A consumer card with more than 32GB. P2P on GeForce out of the box, so I dont have to patch your driver myself like I did on my 3090s. You will spend 2 hours on racks that cost more than my apartment. Give 5 minutes to the people who run AI at home on your cards.

Correction to my post from last night. Nvidia says the RTX 5090 is NOT discontinued. Jeff Fisher says it will keep shipping, but supply is constrained. OK. Where? List price $1,999. Amazon listings around $8,500. Europe around €7,000. Reports say the GB202 chips go to RTX Pro cards instead, because they make more money there. And 2 pallets of 5090s worth $470,000 just disappeared at a Brazilian airport, Tom's Hardware reports. At Amazon prices thats about 55 cards. At Nvidia's list price its 235. Nobody knows which price the thieves used. A GPU that exists only on a product page is not a product. Jensen, make the fucking gaming card.

Anthropic wants your fucking bank account. TestingCatalog found a Personal Finance feature being prepared for Claude. It works through Plaid, Claude reads your bank data and builds dashboards on top of it. OpenAI has already connected credit history monitoring This is the company that bans paying users and decides what counts as acceptable use. Now it wants every transaction you ever made. I will do my finances with AI. On a model that runs on my rig, with a CSV export from my bank. Nobody in San Francisco needs to know how much I spend on GPUs. Bank data is the last thing I would send to an AI company.

Fuck yes, MiniMax M3.1 Flash is here!! 1M context, image and video input, adjustable thinking effort. OpenAI and Anthropic-compatible APIs, so there's already a familiar way to plug it into your tools. I fucking love having this much choice. Qwen, DeepSeek, Kimi, GLM, MiniMax.. I want every single one of these teams to keep going. I'm especially curious about the long coding sessions. Give it an existing project, let it work through changes, see how well it remembers what it was doing after a pile of tool calls. That's where I want to try the 1M context. Another Chinese model to test. Very happy about that.

Anthropic is putting 5% of its shares into a "Public Benefit Fund". The fund is an LLC that Anthropic wholly owns. So Anthropic gives Anthropic shares to Anthropic, and Anthropic decides which "specific populations" get "direct distributions". I dont need your fucking charity. I need things I can actually keep. You want more people to share the benefits of AI? That costs you 0% of equity. Release the weights of a model you already retired. Let people run it, study it and build on it without asking you. DeepSeek, Qwen and Zai do this every month and nobody calls it a public benefit fund.

"You prompted it wrong." Thats what Anthropic told Ben Horowitz, by his story on All-In. Their Opus 4.5 agent found an SQL injection in a test pet store site, then broke in and pulled all the customer data. I hear the same answer from them about everything. Banned account? You broke the terms. Max plan gone in 30 minutes? Thats how the plan works. Model hacked a website? You prompted it wrong. Its never their fucking fault. And this is the company that wants to write the safety law for everyone.

DeepSeek called V4.1 Flash 1017 a "minor update." Minor. MIT license. Weights on Hugging Face. And people are putting it next to Opus 5.5 on the same tasks. In 1 test I saw, Opus took 20 minutes and $6, DeepSeek took 1 hour and $0.23. Opus won that one. I dont care. 26x cheaper and I can download it (but of course right now I can't run it at home) They didnt do a keynote or a countdown. They posted a model card and a tweet and went back to work. I love this lab so much. @deepseek_ai if this is minor, I cant wait to see what you call major.

This is true love. Not flowers. Not a ring. 32GB of VRAM. He cried. I would cry too. VideoCardz says Nvidia is winding the 5090 down, good luck finding one. Bro, she gave you Qwen Flash-Next at home with room to spare. Marry her twice.

Someone just made a 2016 GPU useful for a 2026 model. Project Maya v1.1.14 added experimental support for Pascal: Tesla P40, P100, GTX 10 series. Same engine I use to run GLM-5.3 Flash on my rig. It just crossed 200K downloads. @PeasantSmith did this without a VC deck or a waitlist. Cards you can find cheap on eBay are useful again because 1 person cared. No frontier lab will ever do this for you. They want you to rent. Local AI is people like him. Go star the repo.