You live on floor 5000 of an endless megastructure. It's snowing. You want chips. The store is one floor up, over a bottomless drop. A tiny first-person game in one HTML file. EN / RU / 中文 claude.ai/artifact/SYh498jjW… artem-x-meta.github.io/etazh…

🚀 Meet Qwen3.8-Omni-Flash, Qwen's first omni-modal model built around agentic capabilities! Native audio-video understanding, reasoning, and tool use come together in one model: understand the content, plan the task, execute with tools, and deliver the result. Highlights: 🥳 - Audio-video intelligence that gets things done: jointly reason over what's seen and heard, and orchestrate tools across long workflows to auto-edit vlogs, translate short videos, and turn movies into recaps. - A major leap: approaching Gemini 3.8 Flash in audio-video capabilities; +19.5 points on average in agent performance across WildClawBench-MM & UniClawBench. - 1M-token context with agentic perception: actively explore long videos and locate key moments with higher accuracy, using 51.8% fewer tokens than static understanding on OmniVideoBench. Video input costs are reduced by about 89% compared with Qwen3.5-Omni-Plus, making long-form audio-video understanding and agentic workflows more affordable than ever. To help you build apps around Omni, we're also open-sourcing Qwen-MM-Plugins and Qwen-Live Harness! 🛠️ We can't wait to see what you build with Qwen3.8-Omni-Flash! 👀 - Blog: qwen.ai/blog?id=qwen3.8-omni… - Qwencloud: qwencloud.com/models/qwen3.8… - Qwen Studio: chat.qwen.ai/ - API: alibabacloud.com/help/en/mod… - Qwen-MM-Plugins: github.com/QwenLM/Qwen-MM-Pl… - Qwen-Live Harness: coming soon github.com/QwenLM/Qwen-Live-…

Новая концепция kinda приложения из коробки: ставится не апка, а запускается промпт/последовательность промптов, создающих апку Ввиду случайностей и юзерских особенностей апка получается уникальной, но при этом сконнекченной с сетью остальных через кредсы заданные соответствующим образом Мнения?

長いエロ動画から、オーガズムを含む10秒間を見つけて、ComfyUIの動画生成用参照素材として切り出す仕組みを、ここまでほぼQwen3.8 27Bと作ってきた。 手作業で大量の動画を確認するのは大変だし、発情してしまう。そこで、AIに動画を見せて「どこがオーガズムか」と聞いても、まだ正確には判断できない。 試しにとあるAVモデルさんのエロ動画の 音声・映像の動き・動きのフリーズのスコアを眺めてたら、 オーガズムの瞬間の特徴として次のような傾向があった A. 100〜1500Hzの喘ぎ声→その後の急減衰 B. フレーム間の映像変化量急増 C. freeze、コマ止め、ディゾルブなどのソフト遷移 試しにこれらから条件を決めたスクリプトを書かせて、オーガズム候補の位置を探したところ、結構ビクビクなってる感じのところが溜めれたからまあ良いんかなと。 この後は、それを自動化して、24fps・H3規定フレーム数へ整形して、そのままComfyUIへ投入できるようにする。 もう少しデータが集まったら、絶頂の瞬間とそうでない瞬間の数値データを学習させ、CatBoostくらいの軽量な機械学習で自動的にカットを切り出せるようにしたい。 ただ、オーガズム時の声や動きは人によって異なる。最終的にはAVモデルごとに学習モデルを作るか、個人差を補正する仕組みが必要になるとは思う。そこまではQwen3.8 27B でいけたら面白い。 グラフのGTが絶頂の瞬間。 A B C 全部で絶頂かと言われるとそうでもない。 この辺のパターンが人間だと良くわからない→分類器作ろ。 この発想は、投資の売買シグナルの選別によく似てると思う。

I spent 1351 hours of model time, ran 17555 generations, and burned 91.4M reasoning tokens. 300 benchmarked tasks. 9 versions of the exact same Qwen3.8 27B. Around $100 on Blackwell GPU rental. Just one single question: which quant actually degrades quality? The results will surprise you. 🧵Mega-thread, read till the end, the takeaways are right there

Right now, I'm launching an experiment for 200 inference hours of Qwen3.8 27B. I will run 300 tasks in low, medium, and xhigh modes on FP8, AWQ, UD-Q6_K_M, UD-Q4_K_XL, UD-Q3_K_XL, UD-Q2_K_XL, and UD-IQ1_M. Then I'll compare each version against the BF16 baseline. Wish me luck. And may God help me! 🙏

Nobody ever shows you the log. Here it is: 687 steps of a local Qwen3.8 27B building a first person shooter in DeepSeek Harness, scrolled end to end. It takes a while. 87.9M tokens went in, 822k came out. 5 hours 11 minutes of model time, 58 minutes of tool calls. All of it on two used 3090s in my apartment, none of it on anyone elses hardware. Watch for the spots where I stop typing descriptions and just paste pictures. A screenshot of the gun I wanted in my hands. A screenshot of the crosshair I wanted on screen. The model looked at both and built what I meant. Thats a 27B doing art direction with its eyes, on my own cards. The rest of the log is not pretty and thats the part I want you to see. It writes a file, runs it, reads its own error, goes back and fixes it. Six hours of that while I slept. Everyone posts the finished demo, almost nobody posts the conversation behind it, and the conversation is where you find out whether a model holds a task or just talks about it. This one held it for 687 steps without me stepping in. Local models are past the demo stage. Watch the scroll and tell me you saw this coming 18 months ago.

Qwen3.8 27B wrote a playable first person shooter start to finish on 2 used 3090s in my apartment. No cloud. No API key. No subscription. 687 steps. 87.9M tokens in, 822k out. 5 hours 11 minutes of model time plus 58 minutes of tool calls. The agent loop never broke once. And it plays. Enemies spawn and push you, the gun kicks, shadows stretch across the whole block while the sun goes down behind the towers. I sat there clearing waves instead of grading the output. Same shooter prompt I threw at the frontier models a few weeks ago. That time the tokens went to somebody elses datacenter. This time nothing left the flat. 87.9 million tokens through my own cards. On an API that run has a price tag. Here it has an electricity bill. 60 tok/s all the way through. Slower than frontier, and it stops mattering when the thing works through the night while you sleep. Local models were a toy 18 months ago. This one finished a game.