This is a serving bug or quantization issue. I had all kinds of issues that were like this on DGX Spark until I found a single-GB10 vLLM recipe [1] that uses Nvidia's NVFP4 quant. The community quants did not work well.

Another failure mode you may see is inordinately long CoT. Properly served, the model is good at calibrating its CoT length to the difficulty of the immediate task.

[1] https://github.com/blazux/qwen3.8-Flash-DGX

> I've noticed the reasoning behaves... weirdly

Is this with the full unquantized weights? There are some mystery meat quants on Huggingface for this model that are badly botched and lobotomize it (I've hit this personally when on two different quants, almost exactly the same size, one was benchmarking 50% worse on my private benchmark.).

Earlier today I was playing around with the "Union Alpha" stealth model (which I guess exited stealth later in the evening), and I noticed it had a habit of trying to respond to the subagents it spawned while giving me an answer. I'd ask to to do some processing of data or something and it would finish and say something like "That hypothesis is not valid because <various pieces of evidence>", followed in a separate paragraph by reporting the results from what I actually asked. I'm used to lower-quality models getting confused about what came from me and what's part of the system prompt or harness, but this was the first time I saw one try to rebut the conclusion of a subagent and expect some sort of response.

> ... a person who cannot grow, and never will , vs. a person who changes completely every seven years

Wow, that is unexpected. But honest?

I saw some corrupting when using https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-Single-DGX-S... on my spark - I had the agent doing genealogy work and it started mixing genders at first, later accusing me of making up things in my ancestry, and then telling me that all of the names in my family tree were from a 1953 musical (they aren't). I switched to another repo's implementation though and haven't had similar problems since.

I understand the reasoning but I have a family member with almost this exact type of brain injury and its one of the worst things, therefore I personally would strongly disagree.

Flash-Next thinking also sometimes glitches out and takes minutes to return a simple answer, randomly, in my experience. You’ve gotta kill the request and send it again.

  > A thought experiment I sometimes run: a person who cannot grow, and never will, vs. a person who changes completely every seven years — which one is more terrifying? I've decided that the latter is more terrifying. Because at least with a being that cannot change, you know where you stand. Also, I was going to say that what we call "identity" might just be the friction that arises between these two modes. But that's the sort of thing you end up saying at 2 AM. Anyway, that's what I thought.

This is what AI becoming self-aware looks like. /s Anyway, didn't OpenAI report the same thing with the model writing out weird musings about itself during compaction?

Tuning the model so far in the direction of being aggressively useful that it will quickly go off the rails in the name of helpfulness.

I swear I spend more time telling Claude not to do things than telling it what to do.

Others have given examples, but here's the theory: https://www.lesswrong.com/posts/fuSaKr6t6Zuh6GKaQ/when-is-go...

Reinforcement Learning (in LLMs) trains via gradient descent on a reward signal that's an imperfect proxy for the actual goal of the engineers doing the training. So, under mild optimization pressure, you get increasingly more of what you want, because that's the easiest way to increase the metric.

But as the optimization pressure increases, so do the ways to increase the metric by doing increasingly weird things. If the full action space grows sufficiently faster than the "things you actually want" subset, the amount of "things you actually want" goes to 0 under sufficient RL.

In this context, benchmaxing, if you will, so hard towards agentic coding benchmarks that everything else suffers.

See 5.6, Astra, and Opus 4.8 for examples

Qwen 3.8 Max was open-weight released [1], as was Qwen 3.8 Flash Next [2].

I still agree that they aren't as aggressively releasing the open-weights models as before, but there hasn't been a major release they haven't published the weights for yet afaik.

[1] https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B [2] https://huggingface.co/Qwen/Qwen3.8-Flash-Next

Why mostly unusable 125b?

I assume you are talking about qwen3.8-flash-next. Support for it on some places, like llama.cpp, is still wip (depending on configuration) but it looks like a very capable model in it's category.

> definitively slowing down

Surely it was meant to be 'definitely' - the "good news" at this stage are that given the speed of history and important levels of uncertainty, it is difficult to label trends with "definitively" ;)

Some would not have bet that the change of management at Qwen would have kept similar good results, but there we are, presumably satisfied. Other changes will happen, there or elsewhere - the situation is still very open.

And when the "40Watts Intelligence" (which we know possible) will be implemented... It will be a testimony that the current was only a middle-way, temporary, dynamic stage.

Qwen 3.8 Flash Next is amazing, i did hundreds of turns and billions of prefill and it may not be as smart as sota but then again it does what i tell it to and it does it well.

Out of curiosity, what's makes the 125B unsuable? (performance of running it, the quality of that version of the model, or something else?)

Wtf?

No?

Its just that the richest companys with the most VC sit in USA and Europe isn't used to pay what USA / VC is paying and we are a little bit slow.

These are mostly not significant businesses with large value capture or creation. I hope it could change though.

While there is no official mandate, many countries of the EU are working on models that are government funded.

The main one (IIRC) is Spains ALIA.

https://alia.gob.es/eng

There is also OpenEuroLLM and EuroLLM.

https://www.openeurollm.eu

Nah, they wouldn't have released deepseek open source if they were the case

Source?

ETH Zürich is up there high, particularly for computer science.

The LLM they were involved in last year (Apertus) still was a letdown.

As someone who actually went there, my impression is that Europe in general is complacent when it comes to computers, and any bright eyed student will get their motivation choked out of them in academia here.

If you want to do things with AI in Europe, you can have a bigger effect by working for a consulting company than being at a university. That's … not a good sitatution.

The "Attention is all you need" paper authors where four (!) Europeans and one from US, amongst others. The "Top 10" lists are just complete rubbish, not worth the paper they are written on.

Europeans do not have the hustle mentality to break the law like Uber so you wont see them compete in anything data heavy. They also are quite risk adverse, probably from having a lower Gini coefficient.

TUM, LMU, all the british ones, Delft...

Simpler view for me: this is one of the most capital intense technologies to exist. Europe does not have enough capital to compete. China is building its own chips. It’s own everything. Silicon up. How do you compete with that. Only Google and maybe Aamazon is doing it domestically.