Qwen rumored to push 5-10T params with recursive self-improvement, plus Yunqi lineup
lxfater · x · 2026-09-22
Unverified claims mixed with conference news.
- The post claims Alibaba announced Qwen moving toward 5-10 trillion parameters, using a new method called RSI (recursive self-improvement): Qwen3.8-Max allegedly ran 33 rounds of self-directed training over a month, lifting benchmark scores from 40 to 45 with no human intervention.
- RSI is also said to be applied to inference optimization (96% single-instance throughput gain adapting a new GPU) and chip design (60+ hours of autonomous runs, 42% smaller die area, 59.5% lower power).
- Yunqi conference releases: Qwen3.8-Omni (omnimodal), Qwen-Image-3.1, Wan3.0 (top of two video leaderboards), Qwen3.8-LiveTranslate (sub-2.5s interpretation latency), HappyOyster 2.0 Preview world model, Happy Shrimp 1.1 music model.
- Qwen-Audio-3.1 covers ASR/TTS/realtime speech; TTS-Next generates voice, SFX, and ambience in one pass, up to 2 minutes, 14 languages and 29 dialects, #1 on HF TTS leaderboard.
- Qwen3.8-Flash cuts training cost 90%; Qwen3.8-27B runs on consumer GPUs, dubbed "local Opus4.6."
Related event: Qwen4 in Training, Aims for 5-10 Trillion Parameters Next(2 posts)→
More from Models
- Anthropic's 'Constitution' Is Just RLAIF, Argues Researcher — Its Real Benefit Is Cutting Human Labeling — WillRinehart · 2026-09-22
- Qwen4 training now: Alibaba teases Max/Flash/Plus variants, 5-10T params for Qwen4.5-5 — ChrisGPT · 2026-09-22
- Fable can interrupt itself: answer your new prompt, then resume the first — gleech · 2026-09-22
- Did OpenAI Solve the Wrong Navier-Stokes Problem? Experts Cry Loophole — joshgans · 2026-09-22
- Qwen4 family incoming: Max, Flash, Plus and the 27B for local runs — kimmonismus · 2026-09-22
- Google ships Gemini 3.8 Live with background reasoning during voice calls — emmanuelvivier · 2026-09-22