Alibaba releases Qwen3.8-Max, a 2.4 trillion parameter model, with open weights coming next week
- Alibaba released Qwen3.8-Max, a 2.4 trillion parameter model (95B active) built on the Qwen 3.5 architecture, and will open source its weights next week, the first time for a Qwen-Max-class model.
- In a 10+ day autonomous coding test, Qwen3.8-Max built the oh-my-cli project from scratch using a self-evolving harness, producing 265 commits, 127 PRs, and 151 issues over roughly 16 days with no human help.
- Given only a research paper ("Unified Data Selection for LLM Reasoning") and GPUs, the model worked alone for about 125 hours, wrote ~7,600 lines of code, took 1,100+ actions, ran 33 rounds of GPU training, and reproduced the paper's six main findings including a +7.7% gain on AIME24 over random data selection.
- Alibaba also plans to release Qwen3.8-27B as open weights next week, positioned as a successor to the widely used local model Qwen3.6-27B.
- QwenCloud pricing for Qwen3.8-Max is $2 per million input tokens and $6 per million output tokens, with $0.25 per million tokens for implicit caching, and it supports a configurable reasoning_effort parameter (xhigh, medium, low).
Hacker News 의견들
Wait, are these Qwen models still open weight? The article literally says right in the second sentence they're open sourcing weights next week, first time ever for a Max-class model.
Funny how their promo video shows people doing hobbies while AI does their job. Is Alibaba leading by example by cutting work hours while keeping pay the same? Didn't think so.
Honestly I'm busier with AI than before. I used to have 10 tasks taking an hour each, now I have 100 tasks taking 10 minutes each. Same time spent, maybe more stressful, just more output.
The percentage of people who actually need to work keeps dropping from aging and automation. Already 38% of Americans 16+ don't work and aren't looking, and they're not starving. Would've looked like utopia to someone in the 1800s.
Confused about the timeline here. There was already a Qwen3.8-Max-Preview announced on July 19th that I tested (drew a pelican with it), so what exactly is being released today?
They're releasing the full Max now, not the preview. Two weeks ago was preview, this is the official checkpoint after more RL training presumably.
Just tried the new model with my pelican test, took 11 minutes and it forgot the wheels lol. Cost me about 17 cents at e also announced Qwen3.8-27B open weights coming next week too. If it beats Qwen3.6-27B that'd be huge since nothing else touches that size class right now without being benchmark-gamed.
I'm still running fine-tuned Qwen3.6-27B deployments with good results, just need a bit more intelligence bump. Hoping 3.8 delivers.
Qwen3.6-35B is my daily driver, it's what got me to cancel my Claude subscription back in April. Best local model I've tried by far.
Pricing is 6 per million output, plus $0.25 for implicit caching. Was hoping it'd be cheaper than that honestly, I've gotten used to Deepseek's near-free pricing.
This 'self-evolves through feedback loops' language, does this just mean they distilled Claude? Sounds exactly like what Claude Code already does.
It's a meaningless claim really. Models have always been able to iterate like this, RL just reinforces it by rewarding exploration. What actually matters is how long it runs without human intervention, not that it's possible at all.
We need an actual benchmark for this self-evolution stuff to see which models can really create recursive improving solutions versus just marketing copy.
Once OpenAI and Anthropic go public, every one of these competitor announcements becomes a reliable sell signal for their stock.
The fear mongering from Altman and Dario about needing to ban foreign models for national security reads like they just want to corner the US market and keep profits, since China's already proven LLMs are a commodity.