跳到正文
Hacker News:AI 熱帖·· 4 天前AI 評分45

Qwen4-Max 登上 Hacker News

Qwen4-Max

AI 導讀

Qwen4-Max 登上 Hacker News 後,留言討論模型效率、品質及能力發展。有人期待「Flash」和 27B 版本;另有用戶稱 Qwen 3.8 每秒約生成 30 個模型 token,Opus 則接近 80 個。

正文

Image 1Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Qwen4-Max (qwen.ai) 18 points by ilreb39 minutes ago | hide | past | favorite | 3 comments help Image 2seizethecheese5 minutes ago | next[–] I hope they are able to serve 4 faster. I’ve regretted including Qwen 3.8 in the battle system for opencode [1] we have been building because it scores so low. And it’s a shame because the only reason it scores so low is that it’s slooooow. (Like 30 tokens per second vs opus near 80.) If you exclude voting, Qwen 3.8 is near the top. (We allow early voting contrary to Arena.ai because I think it makes sense for agents, where it doesn’t for chat.) Also, it would be ideal if they served this with third parties. Part of the slowness with 3.8 is just not a lot of providers. Their open source version is much faster, but we couldn’t include it because it doesn’t handle images (another own goal by Alibaba). [1]: https://agentduel.com/leaderboard reply Image 3dan_thinks25 minutes ago | prev | next[–] This is exciting. The continued focus on efficiency alongside quality is encouraging - it makes me hopeful to see what smaller models will become capable of. Looking forward to the "Flash" and 27B versions. reply Image 4bbor12 minutes ago | prev[–] Ah yes, an open-weights, soon to be abliterated model that beats Opus 5.5 in checks notes bioinformatics and "legal research". Some part of me is excited for the coming catastrophe(s), because it'll end the frustrating last ~12mo of "those Anthropic jerks just want regulation so they can corner the market on artificial minds wrought from sand!". Another part of me knows that; A) Rapid epistemic decompression of this scale will probably lead to some truly nasty, illogical sentiments, not to mention other pathologies (deep depression, compulsive anxiety, etc.). Maybe some will reasses their beliefs after some sober reflection, but a minotiry, I expect. B) Trading hundreds, thousands, millions, or billions of lives so that I feel a little less terrified and alone isn't exactly compelling utilitarian calculus... Anyway, to the tech: another impressive jump for Alibaba, catching Moonshot overall with some interesting niche improvements and a massive, gaping flaw in the poor TerminalBench scores. I'm probably just biased as a SWE, though. I can't help but feel like Ziphu will be the ones to really blow our socks off next, given how long it's been since GLM-5.3[1] and the vague rumours about the next one (5.5...?). The Chinese labs' leaps to catch up to Anthropic & OAI from the strange, long-forgotten past of September 2026 were clearly due to a tripartite distillation campaign, so expansive and rushed that they accidently leaked tons of CCP state secrets.[2] Has anyone heard any rumours about if that's been continuing after the crackdown? Anthropic has been pretty damn aggressive with catching distillation, so my personal guess is that this is just honest improvements. ...which would mean that at least 2 of the 4 Chinese labs are ~3-5 months from the RSI possibilities Anthropic is trying to sound the alarm about today, if that. Jesus fucking christ. We need to rally around international protections ASAP, cause something tells me at least one of the four won't quite so ready to "pace" themselves in the current climate. --- [1]: https://www.anthropic.com/research/glm-5-3-and-the-spread-of... [2]: https://www.anthropic.com/threat-intelligence-report-septemb... reply Image 5 Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact Search:

來源:Hacker News:AI 熱帖 · news.ycombinator.com