Qwen 3.8 27B (dense) running on a single RTX 4090 (24GB VRAM) at 65 tokens/sec decode with MTP!
260,000 context window or 65 tokens/sec decode with native MTP. The API cartel should be terrified. We are officially running frontier tier agentic AI (benchmarks comparable to
Qwen 3.6-27B is 14x smaller than Alibaba's flagship and BEATS it at coding.
A 27B model just outperformed a 397B model. And it runs on your own computer. 🤯 The numbers: → 77.2% on SWE-bench. It fixes real GitHub
Qwen 3.8 Max is 50% off on Venice for a limited time, in collaboration with Alibaba_Qwen.
Qwen 3.8 27b (output via Lmstudio chat)
Singleton From ‘Please make a Google dinosaur game.’ Resulting video. Advantages: Excellent performance not seen in models smaller than local 30B. Disadvantage: Very large inference bubble and resulting increased operation time. (Token 61408 used, 37.9Tok/s: 3090, 21 minutes 29 seconds) Alibaba_Qwen QwenDevs…
You have reached the end of the archive
All of qwen38