Qwen 3.8

I’m a developer from China. So, is this what Hacker News is all about? Whenever a model comes from China, the comments section stops discussing its technical architecture, optimisation points or real-world performance, and instead starts going on about politics, human rights and all that rubbish? To be honest, we Chinese IT professionals possess a genuine geek spirit. That’s why you’re lagging behind in open-source competitions—it’s got nothing to do with politics; you’ve simply lost sight of your original aspirations.

(I can't draw a pelican for this one because Alibaba Cloud have flagged my email address and won't let me pay them for access. So I'm waiting for the open weights release, or for the new model to show up on OpenRouter.)

Without a doubt, the Chinese models are going to overtake the American ones. In my opinion, this will happen this year!

I assume that this announcement has been prompted by that of Moonshot AI, which has just announced a 2.8T parameter open-weights LLM, Kimi K3, to be published on Huggingface by 27 July.

Now the response of Alibaba is that they will also publish soon a big open weights LLM, the 2.4T parameter Qwen 3.8.

I wonder if Alibaba has always planned to make this big LLM open weights, or they have chosen to do this now, to better compete with Moonshot AI.

In any case, from this competition in LLMs, we win.

Deepseek 4 "final" version is imminent as well.

Will probably be at Opus 4.8 level, and I find it pretty big deal because of Deepseek price...

in my experience of 1 month daily use, Qwen 3.7 Pro is just unusable. wastes too much time, goes off track, useless stuck loops, cannot debug at all. Deepseek V4 Pro is night-and-day compare to Qwen. actually Qwen models seems the worst SWE experience so far. and it is super expensive compare to Deepseek. cannot delegate anything to it, cannot use it real-time low-level tasks either. totally unusable.

Qwen is the most censored of the Chinese models in my testing, which makes me wonder in what other ways it is compromised. Open weights doesn't really reveal what's in there. And, in my tests, existing Qwen models are not at the pareto frontier of any metric; DeepSeek V4 Pro is better, faster, and much cheaper than Qwen 3.7 Max. (DeepSeek is also among the least censored of the Chinese models.)

I guess we'll see if the "second only to Fable" hype pans out. In my limited experience with Kimi K3 (I signed up for a month of the $19 plan) it's slower and chews a lot more, so ends up being pretty expensive; one little feature burned through almost the entirety of my five hour limit. The $20 GPT plan is a lot more useful and includes 5.6 Sol, which is fast and token-efficient enough to be quite usable even with the small plan.

I've been using Qwen 3.6 27B with LMStudio, and I was pleasantly surprised with it, although it was a little slow. I found mtplx last night, and it really wasn't an exaggeration to say that it ran the model 2-3x faster which was super impressive.

I'm trying to move to local models as much as I can, and I'm finding that it's becoming more and more practical. Admittedly this is on a $6000 dollar laptop (M5 Max Macbook with the specs maxxed out), so the hardware is still a bit out of reach for most people (the AI industry isn't exactly helping here..), but I'm getting the impression that the future is going to be smaller models with more focused training running locally. The danger of giving all your data to these cloud providers just seems too big to me, and I think they're going to start charging insane amounts when they need to show a profit.

I counted the other day and there were at least 12 different providers with "better than Opus 4.5 performance" on Artificial Analysis, Opus 4.5 being Anthropic's December release that many say kicked off the latest acceleration. Which is totally insane competition, particularly given how low switching costs. I personally think that Opus 4.5 level performance is sufficient for most apps and usecases, as they get deployed.

I have been using the docling software for PDF and XLSX reading/viewing/comprehension by my local qwen3.6. Claude was telling me that docling includes within it a very small LLM model just to assist with what is basically "super OCR". We are definitely in this era of ultra-mega-huge frontier models and super-tiny-micro models, all finding their uses.

On an aside I tested docling by giving it the Ronin TTRPG rulebook PDF and it did an astounding job of converting it all into .txt and .md. Given the wild graphic design of the book that's pretty astounding. Next we'll see if it can read the other Borg books like Mörk Borg and Cy Borg, which are also famous for their messy information.

Bring it on! Hoping that they release smaller sizes of Qwen3.8. I use the 35B MoE and 27B dense models locally and most of the time I don’t need to reach out to Claude. Extremely useful specially when requests include sensitive and/or personal data

For those trying to get it to work in OpenCode with a Qwen Cloud Token Plan, this is what worked for me. Note that I've just matched Qwen 3.7 Max for the limits as I don't know exactly what they are.

  "provider": {
    "alibaba-token-plan": {
      "models": {
        "qwen3.8-max-preview": {
          "limit": {
            "context": 1048576,
            "output": 65536
          },
          "modalities": {
            "input": [
              "text"
            ],
            "output": [
              "text"
            ]
          },
          "name": "Qwen3.8 Max Preview"
        }
      }
    }
  }

I really like Qwen, even the Q2KP quants of 3.6 27B have genuinely impressive local performance on a 24GB card. It has been good enough that I am happily giving them $60 right now to try this instead of waiting to try a slightly lesser version locally.

Was there ever an explanation for why we never got the weights of 3.7? I would like sourced quotes and not weird/cringe accusative speculation about distillation, or your take on The Big D.

Qwen has set an excellent track record for architecting and releasing open-weight models that consumer-grade devices can run. What is needed the most right now is something similar to Bonsai 27B, with a modest memory footprint, but faster and more capable. On-device models can make up for intelligence by being faster, thinking longer, or doing more quick iteration rounds.

Things are heating up in China.

Looking forward to see what Antrophic and OpenAI does next.

I'm haven't found an announcement page, but there's a banner on the website announcing Qwen 3.8 and redirecting to this page.

Looks like they're previewing the model only on their subscription plan.

> "compatible" instead of "comparable."

It boggles my mind how you can train a frontier model but not write a tweet without an obvious typo.

The "second only to Fable 5" comment is pretty telling here. I remember early on when a lot of naysayers were saying that Fable was barely an improvement on Opus. Like it or not, Anthropic have a genuine moat right now with that model, provided they continue to allow people to use it. It will be genuinely exciting when an open model is able to beat it.

> You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork. Be among the very first to try it out.

You can also try it out on Qwen chat, Its free.

> With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5.

That's a massive model!

The shift from "value" models to "intelligent, huge and slow" models coming from China is an interesting change in strategy.

My main issue with GLM 5.2 and Kimi 3 is that they're extremely token hungry and thus feel slow(er) to use.

The title is misleading. The link led to a pricing page/token plan and not about the new QWen 3.8 model.

I remember when they released Qwen 3.7 Plus and Max. These models behaved way different from all prior models, it became too verbose. It wrote multiple paragraphs just to answer my prompt instead of the usual concise and direct way responding to me. I didn't like that at all, and I know Gemini also had this behaviour with with the Flash series until I managed to reduce it a bit with personal instructions (in the settings on Gemini website).

I haven't tried Qwen 3.8 Max yet, looking forward to it. My hope is that its way less verbose. Another thing I experience with the Qwen models is that I do not trust their benchmark scores at all. Have anyone played with Qwen 3.8 Max and can share their experience? Which model it come close to? Sonnet 5? GLm-5? DS V4 Pro? Flash? Gemini 3.5 Flash?

Now there are not one, but two incredibly powerful open LLMs. I think this level of capability makes general prioritization / high level decision making doable with the right harness, and now everyone has hard-to-interrupt access to them (since these are open weights and someone in the world is going to run them). This world is going to get really weird soon, in both good and bad ways...

I do hope they provide optimized A3B quants--that's been the sweet spot for usable local inference for me, at least.

What is a "credit" and how does it translate to tokens for the different models?

Why did Qwen stop producing open models? They've gone from building the best open models ~1 year ago to producing like the 10th-best closed models. I don't understand this pivot at all.

Edit: I saw online they do in fact plan to release this openly at some point – x.com/Alibaba_Qwen/status/2078759124914098291

Wait.. the Qwen Max models have never been open-weight. But it sure sound like that's what they intend now?

"Qwen3.8 is launching and going open-weight soon! With a massive 2.4T parameters..."

I did some comparisons to Kimi K3... and found the best is combining the two! Model diversity is powerful.

https://trilogyai.substack.com/p/qwen-38-max-benchmark-how-i...

Always nice to see more open-weights in the heavy model class. I can only hope this trend continues, causing OpenAI and Anthropic to crash and burn.

Has there been any news on open weighting Qwen Image 2.0 and WAN?

We are spoiled in the LLM segment, but I would love to see an open source competitor to Flux.2, etc.

Off topic, but comment threads like this, where the early comments rabbit hole into something I didn’t come to read, really make me wish HN had collapsible comment controls. My phone screen nearly melted from the insane amount of swiping/scrolling to get to the part of the discussion about the actual LLM.

Does anyone have the privacy policy of their token plan available? Want to check if they retain/train on inputs/outputs.

The few tests I ran were by no means comprehensive, but while kimi felt like the real deal qwen seems a bit of a benchmark princess.

Qwen is so much better than GLM or Kimi that this makes me genuinely excited!!

Does anyone know if they intend on releasing open source/weights variants for 3.8 or whether 3.6 was the last model they are/were doing that for?

The only problem I had with Qwen, fine tuning on Colab, it takes 31 t/s while Gemma 4 is around 9 t/s, otherwise, one of best local LLM

With 3.6 27b, I just stopped changing local models and started tinkering with things on top (like mem0). Feels genuinely useful and more than a toy

Looking forward for a 20ish billion parameter version

These models are great, but what's the potential use? No small entity can run them.

If 3.8 max goes open weight, what are the odds they retroactively open weight the earlier releases?

Open Source Era is comming

It feels like an inflection point of lost US leadership in technology? A year plus ago you would say while China led in green energy and manufacturing, at least the US was ahead in software - as demonstrated by the state of US AI models.

We could point at a lot of factors on the US side. From political paralysis / head-in-the-sand attitudes towards emerging tech like green energy. To something of disdain for workers that will be impacted by AI (creating a backlash). To education that continues to lag. Add to this so many other self-inflicted economic wounds from the current administration.

I don't know if its nearly as terminal, as say the UK after WW2. The US is still large, wealthy, and resource rich. Yet at a minimum the triumphalism about US leadership after Trump was elected by the tech elite feels silly in retrospect.

Something I also think about is how much stronger The West overall would be if instead of antagonizing allies, there was a single ecosystem working closer together.

in case someone wants to speculate in why chinese labs open source their models: https://try.works/why-chinese-ai-labs-went-open-and-will-rem...

If it offers more than opencode go, the entry plan looks enticing

I predict that no one will use this and everyone will use Kimi K3.

Who is behind this site? Is this another frontend to Alibaba or a reseller in Singapore?

Go China, screw America*

*within the scope of open models only

can i get "code instruct" version of this? i want 7B and 14B to launch on my hardware.

interesting, when will this race end?

So are locally-runnable models frozen at Qwen 3.6 now :/