> audio-visual performance close to Gemini 3.8 Flash and overall audio performance that exceeds Gemini 3.8 Flash
Wow crazy if true. I think Gemini's audio capability and multi language was the "selling point" for a lot of people. Other capability also matches or exceeds 3.8 Flash.
They also made a new harness but github link seems to 404.
3.8 Max is the most “grounded” model I think - talks generally normal, doesn’t go crazy and start doing things (I see you Gemini), has good design choices and isn’t overly nitpicky. But god it’s slow. And only available from Alibaba. Their token plan is stingy too. If I had to pick the “old reliable boring” LLM, a modern Claude 4.5 if you will, Qwen is my choice. Hopefully they don’t RL it to oblivion.
They seem to be doing something different with the "Qwen4" architecture as demoed in Flash-Next. I've noticed the reasoning behaves ... weirdly. Like, really weirdly compared to any model I've ever seen before.
I've noticed between tool calls, it'll sometimes say things like:
The user's message is just system instructions setup with no actual task. There's no question to answer yet. I should acknowledge briefly and wait for the actual request.
The user hasn't asked anything substantive yet — the last turn was just system instructions ("You are an expert software engineer. Helps user to solve problems."). My previous response was a brief acknowledgment. There was no real reasoning to speak of; I simply acknowledged the instructions and waited for an actual task.
【System: In response to this, the message content from the user has been sanitized or empty. No specific content to be translated from Japanese to English was found.】
These don't clearly reflect ... anything, and it keeps performing tool calls correctly anyway. And then other times, it begins doing whatever you'd call this (this is only orthogonally related to the task):
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 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.
I've experienced the same, IIRC also using NVIDIA's NVFP4 quant. Also just decided to ignore because it didn't seem to cause any real issues. I figure it might be a training thing, since the hallucinated user messages seem to occur immediately after tool calls or when it's checking its work.
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.).
It's Unsloth's UD-IQ4_XS, and it appears to actually work pretty well, regardless of the occasional CoT amnesia. Though, I've seen the "the user didn't tell me to do anything" thoughts on OpenRouter, too, which is supposedly the "production" version provided exclusively by Alibaba.
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.
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.
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.
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?
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.
I think we are starting be on that territory that regular software development is suffering, current models are great for benchmarks and one-shots but in daily development models are too eager and try to force patterns like excessive tests in every turn.
Curious if or when we'll see the Qwen4 series, one thing I love with Qwen is it comes a much larger range of sizes so I can experiment which extremely small llms.
And maybe Qwen4 won't be released, they only release Qwen3.8 27B (and a mostly unusable 125B). There are definitively slowing down open weight release.
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.
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.
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.
No, it is actually very good. Qwen Flash 3.8 Next is fine. But you need ~128GB of RAM to get it going and not a lot of people have that or can serve it very quickly. I have been running it on an old gaming system around 25 t/s to do overnight work and it is very strong, even at 3 bit quant.
Why can the Chinese build models and Europe cannot? The algorithms behind this stuff are not that complicated, are they? Is it the cost of energy? The illegality and difficulty of obtaining all the data in the world? Lack of capital to start moonshot labs? Lack of optimism?
The Chinese just seem to have an ability to get it done without anywhere near the GPUs of the US and Europe can buy these GPUs.
I think relying on the US and China for AI is probably not ideal? For example I think Qwen have not released the Omni models as open weights in the past, it’d be good to know if they’re doing this here?
In Europe AI is seen as evil that must be stopped. Unlike in the US, AI doomers have Commission well manned. Ursula addressed this in her State of Union speech yesterday. The EU rather chooses consumer protection, even if it means killing AI and maintaining status quo.
Because any AI company would be hit by hate and regulation derived from this, no VC invests in the EU. It's more state and large enterprise investments, good old East Germany style. And historically it has not been that efficient.
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.
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.
I feel people are too focused on US/China and don't pay attention to what is going on in the world.
ASML in the Netherlands for example was the only company in the world that makes EUV lithography machines, which all the major chip companies depend on. China recently reverse engineered their work to create machines since late 2025, but not sold commercially.
Ireland has chip production facilities.
EU might not be in the top 2, but it is not out of the running at all.
If the performances are comparable, and there is no evidence it's not.
in/out ($) Gemini : 1.5 / 9.0 | Qwen 3.8: 0.15 / 0.47
That is a massive cost reduction.
Refs: https://www.alibabacloud.com/help/en/model-studio/model-pric... https://runware.ai/gemini-omni
> audio-visual performance close to Gemini 3.8 Flash and overall audio performance that exceeds Gemini 3.8 Flash
Wow crazy if true. I think Gemini's audio capability and multi language was the "selling point" for a lot of people. Other capability also matches or exceeds 3.8 Flash.
They also made a new harness but github link seems to 404.
probably their distill target
Please what is flash pro ultra and all these, can they just use semver or something
3.8 Max is the most “grounded” model I think - talks generally normal, doesn’t go crazy and start doing things (I see you Gemini), has good design choices and isn’t overly nitpicky. But god it’s slow. And only available from Alibaba. Their token plan is stingy too. If I had to pick the “old reliable boring” LLM, a modern Claude 4.5 if you will, Qwen is my choice. Hopefully they don’t RL it to oblivion.
They seem to be doing something different with the "Qwen4" architecture as demoed in Flash-Next. I've noticed the reasoning behaves ... weirdly. Like, really weirdly compared to any model I've ever seen before.
I've noticed between tool calls, it'll sometimes say things like:
These don't clearly reflect ... anything, and it keeps performing tool calls correctly anyway. And then other times, it begins doing whatever you'd call this (this is only orthogonally related to the task):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 get the same messages in a MLX Q8 quant - I assumed it was reacting to the harness, but it doesn't seem to cause issues.
I've experienced the same, IIRC also using NVIDIA's NVFP4 quant. Also just decided to ignore because it didn't seem to cause any real issues. I figure it might be a training thing, since the hallucinated user messages seem to occur immediately after tool calls or when it's checking its work.
> 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.).
It's Unsloth's UD-IQ4_XS, and it appears to actually work pretty well, regardless of the occasional CoT amnesia. Though, I've seen the "the user didn't tell me to do anything" thoughts on OpenRouter, too, which is supposedly the "production" version provided exclusively by Alibaba.
Can confirm here as well. Running ilintar/qwen3.8-flash-next-gguf-strix-halo (IQ4) on pwilkin/strix-llama.
I see the same thing with Q8.
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.
Quite the model I found this one to be. Disappointed when the trial ended.-
PS: It would be ground breaking if it turns out to have been using Chinese chips for inference, like Stealth Ox Alpha. Unlikely though.-
> ... a person who cannot grow, and never will , vs. a person who changes completely every seven years
Wow, that is unexpected. But honest?
/* An industry that cannot grow vs an industry that changes completely every seven years */
... or months.-
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.
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.
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.
> RL it to oblivion.
What would that mean in this context?
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.
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.
I guess the agentic coding benchmarks don't have many rewards for stopping and clarifying what the user wants?
> aggressively useful ... in the name of helpfulness
But is that because of training, or can that be (also? mostly?) an effect of the "system prompt"?
In this context, benchmaxing, if you will, so hard towards agentic coding benchmarks that everything else suffers.
I think we are starting be on that territory that regular software development is suffering, current models are great for benchmarks and one-shots but in daily development models are too eager and try to force patterns like excessive tests in every turn.
See 5.6, Astra, and Opus 4.8 for examples
What are they examples of? Opus 4.8 was much better than the infamous 5, and I find Astra generally competent.
Curious if or when we'll see the Qwen4 series, one thing I love with Qwen is it comes a much larger range of sizes so I can experiment which extremely small llms.
I don't think Qwen3.8-Omni-X will ever be released.
The last one was: Qwen3-Omni-30B-A3B https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct
And maybe Qwen4 won't be released, they only release Qwen3.8 27B (and a mostly unusable 125B). There are definitively slowing down open weight release.
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?)
No, it is actually very good. Qwen Flash 3.8 Next is fine. But you need ~128GB of RAM to get it going and not a lot of people have that or can serve it very quickly. I have been running it on an old gaming system around 25 t/s to do overnight work and it is very strong, even at 3 bit quant.
Looks like the harness repo is already removed?
Why can the Chinese build models and Europe cannot? The algorithms behind this stuff are not that complicated, are they? Is it the cost of energy? The illegality and difficulty of obtaining all the data in the world? Lack of capital to start moonshot labs? Lack of optimism?
The Chinese just seem to have an ability to get it done without anywhere near the GPUs of the US and Europe can buy these GPUs.
I think relying on the US and China for AI is probably not ideal? For example I think Qwen have not released the Omni models as open weights in the past, it’d be good to know if they’re doing this here?
In Europe AI is seen as evil that must be stopped. Unlike in the US, AI doomers have Commission well manned. Ursula addressed this in her State of Union speech yesterday. The EU rather chooses consumer protection, even if it means killing AI and maintaining status quo.
Because any AI company would be hit by hate and regulation derived from this, no VC invests in the EU. It's more state and large enterprise investments, good old East Germany style. And historically it has not been that efficient.
Or simply: lack of risk taking appetite.
> and Europe cannot?
Well there is Mistral. The EU is ahead on specialised models than general purpose LLMs.
There is.
- Flux3
- Kyutai (Open source AI lab)
- H Company
- LightOn
- AMD Silo (Finland)
- OpenEuroLLM and EuroLLM
There is probably more, but that's off the top of my head.
These are mostly not significant businesses with large value capture or creation. I hope it could change though.
What would be significant in your opinion?
Black Forest Labs for example has a $4B valuation with half a billion raised so far.
The Chinese were directed to it from their government.
We have no such government with a mandate to do that.
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?
Great engineering schools and good tech companies to train them post graduation.
Europe has none;
The best tech university in Europe when compared to Chinese/US equivalents won't even rank in the top 10.
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.
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.
> How do you compete with that.
I feel people are too focused on US/China and don't pay attention to what is going on in the world.
ASML in the Netherlands for example was the only company in the world that makes EUV lithography machines, which all the major chip companies depend on. China recently reverse engineered their work to create machines since late 2025, but not sold commercially.
Ireland has chip production facilities.
EU might not be in the top 2, but it is not out of the running at all.
Europe has more capital than China, it's just spent on pensions for boomers, refugees and pointless vanity projects
It's well documented that the Chinese models were enabled by training on openapi and anthropics models...
Europe actually follows americas rules, hence they're not doing that.
It's braindead for sure considering how the US treats europe, but it is what it's