10 to 12 w per gpu at idle 30 to 35 w per gpu while it generates, peaks around 55 w never past 60°c on either box 45 tok/s fresh,
Bonsai2, a Qwen 3.8 27B model, running on 12GB VRAM -5 hours, 328k tokens, 2,368 lines GPU sitting in local computers is now a local coding agent.
The build is my octopus invaders prompt, which is open, and i'll leave the link below so you can reproduce it.
I gave Qoder one recurring task: → Track Cursor, Windsurf & GitHub Copilot → Check official updates, X, Reddit & Product Hunt → Find the week’s…
Pushing this further, I loaded NVIDIAAI's Nemotron Lightning into the TensorFold engine on my M5 Max MBP and achieved over 200tps decode on the first…
I asked it to build a live stock market dashboard from scratch. It planned the task, built the interface, opened it in a browser, and verified the…
Using Qwen 3.8 Flash Next Quality model on M3 Ultra! AIME 2026 96,7% Context Benchmarks in progress prose and code 💪
This breakdown covers what MTPLX by Youssofal_ is, how it uses native multi token prediction to speed up local models on Apple Silicon
Bilibili のチャンネル「设计之道」が 2026-09-24 に上げた41分の AI
A "beautiful rendition" of world of warcraft in three.js by Qwen 3.8 27b...
Different quant, backend and drafting make this a configuration comparison, not an engine-only test. Runs:
The model is soo tasteful you can feel it in responses. here are the data and specs: qwen 3.8 flash next, 180b moe, official fp8 45 tok/s fresh, 35…
But as context grew, generation fell from about 15 to 8 tok/s across three harnesses.
The speed is absolutely CRAZY 🤯 Local AI is no longer a compromise.
Ik teste het met Qwen 3.8 27b op mijn laptop! De opdracht: maak een robot met uitwisselbare onderdelen, 4 animaties en exploded view.
Made with Tesseract by Mirage Model - Qwen 3.8 27b Hardware - Single Nvidia 96 RTX 6000 Workstation Edn Rate my edits... today...yikes
You can access it for free and as an open-source project at qoder_ai_ide
1M free tokens to play with models like: - Claude Opus 5.5 - Claude Fable 5.1 - GPT-5.6 Sol - GPT-5.6 Luna - Grok 4.6 - Qwen 3.8 Max - DeepSeek V4…
Het model draaide lokaal, bouwde zelf de website en genereerde de afbeeldingen via api. Ongeveer een half uurtje werk, met resultaat!