Qwen

/

Qwen3.8-2.4T-A95B

like 1.2k

Follow

Qwen 103k

También me decepcioné un poco… pero tampoco creo que sea para darle la gloria a Kimi xd, por lo menos Qwen se concentra mejor en su Output final y no genera tan a menudo caracteres chinos de la nada.

lost all good faith with qwen team....

Well, not great, not horrible... There are lighter and faster K3 and incoming DSV4 Pro; and, with 3.6 27B, Qwen is still undisputed king of the <50 GB models.
I hope that 3.8 27B won't be disappointment.
There is a spark of hope, that they might revive 122B, or even challenge DS V4 Flash 0731 (current king of the 90-200 GB range).

Yes if they remove vision it's trash not usable for webdev at all so for what to use it then? Reading emails? Give me a break from this Ai SLOP.
even Qwen 3.8 27B Q8-XL took 1 million tokens to draw simple rectangle and 5 hrs using with python to specified dimensions with formulas it had to calculate. Without vision it could not have done it at all.

vision is vital to achieve SOTA there's no arguing the lack of vision being a good thing period, no and's, if's or but's simple as that. removing vision is clearly a push for people to use their cloud endpoint so they can make money. they are basically doing the DLC thing that gaming companies learned to do or subs. qwen 3.6 27b is still the king and will be for a while.

If opensource community can't make vision support for 3.8 27B then perhaps Kimi K3, DeepSeek, GLM etc. can release mini model with 30B etc.

How about this is it even real turn any model into small one without distillation?
https://www.youtube.com/watch?v=X7wKc4lcvJc&t=5s

真可惜hf不能给模型打差评

You made a post on X without giving any hint that the released model weights wouldn't actually be the full Qwen 3.8 Max.

Here is what you wrote in the model card:

For users seeking managed, scalable inference without infrastructure maintenance, the official Qwen API service is provided by Qwen Cloud.
In particular, Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools, etc. For more information, please refer to the Qwen3.8-Max Overview.

This is just awful. I can already see us getting Qwen 3.8 27B later this week, only for vision input to be paywalled exclusively on Qwen Cloud!

Don't get me wrong — the model is still useful. It's solid for general tasks, and distilling it into pure chat models is great. But when a model is this strong in Vision-Language (VL) tasks and you simply strip that capability away, you throw away half its core value. And I won't even mention the lack of a native 1M context window.

Kimi K3 was released in full, without any cheap tricks like this. Qwen has really changed. I could somehow overlook a reduced context window, but at a time when 1M context has become the industry standard, doing this just feels wrong.
I get that you need to monetize your work and secure a steady revenue stream. But usually, you only paywalled features that no one else in the industry had achieved yet! You kept 1M context exclusive back when nobody else had it. Now, you're stripping away basic standard capabilities.

The open weights for Muse Spark 1.2 will be released soon, and something tells me it will ship with both vision support and a 1M context window out of the box. I'm not sure why it ranks above Qwen 3.8 Max on the arena.ai text leaderboard, but either way: Muse Spark 1.2 is sitting at #10 on the vision leaderboard, while Qwen 3.8 Max is #2. It'll be great to see a top-tier vision model released fully with massive parameters.

In short, props to Kimi.

You know, it's like a parent promising the whole world an awesome Lego set, only to hand you half the box. You stripped away 3/4 of the model's context memory and blinded it completely. In previous releases like Qwen3.5-397B-A17B, you didn't touch vision support and released it freely...

Well, now we know that Qwen doesn't fully deliver on its hype, unlike Kimi. Let's see what Meta does next.

P.S. To top off the absurdity: check the metadata tag right at the top of this Hugging Face page — it is literally labeled License: qwen3.8-max. If that's not enough, open the actual LICENSE file in this repo: the legal document is titled Qwen3.8-Max License.
So they tagged the repository on Hugging Face as qwen3.8-max and shipped the weights under the official Qwen3.8-Max License, yet turn around in the model card to claim that Qwen3.8-Max features like Vision and 1M context are a separate, cloud-only luxury. You couldn't make this up.

If you don't know what you're talking about, don't try to act smart! Not a single AI model in the world has a true, honest 1M token context window — not Claude 4.8 , not Claude 5 fable, nothing. The actual, true linear context tops out at 128k max. Everything beyond that is just a hack relying on interpolation, RoPE scaling, or sparse attention. Go learn how transformers actually work before whining.The same goes for multimodality. If your hands aren't growing out of your ass, you can easily hook up a vision encoder or whatever you want to Qwen yourself. Even models like Claude feature a vision system slapped on with duct tape using a small separate encoder with a hidden projection matrix (around 1k+ dimensions). It’s not native, it’s just modular.

If you don't know what you're talking about, don't try to act smart! Not a single AI model in the world has a true, honest 1M token context window — not Claude 3.5 Sonnet, not Claude 3 Opus, nothing. The actual, true linear context tops out at 128k max. Everything beyond that is just a hack relying on interpolation, RoPE scaling, or sparse attention. Go learn how transformers actually work before whining.The same goes for multimodality. If your hands aren't growing out of your ass, you can easily hook up a vision encoder or whatever you want to Qwen yourself. Even models like Claude feature a vision system slapped on with duct tape using a small separate encoder with a hidden projection matrix (around 1k+ dimensions). It’s not native, it’s just modular.

Like I'd listen to a troll. Besides, you're probably not even a real person anyway. Look at all your threads — you're the toxicity of this community, and honestly, you should just be banned.

Reported for spam and harassment. Have a nice day.

If you don't know what you're talking about, don't try to act smart! Not a single AI model in the world has a true, honest 1M token context window — not Claude 3.5 Sonnet, not Claude 3 Opus, nothing. The actual, true linear context tops out at 128k max. Everything beyond that is just a hack relying on interpolation, RoPE scaling, or sparse attention. Go learn how transformers actually work before whining.The same goes for multimodality. If your hands aren't growing out of your ass, you can easily hook up a vision encoder or whatever you want to Qwen yourself. Even models like Claude feature a vision system slapped on with duct tape using a small separate encoder with a hidden projection matrix (around 1k+ dimensions). It’s not native, it’s just modular.

Like I'd listen to a troll. Besides, you're probably not even a real person anyway. Look at all your threads — you're the toxicity of this community, and honestly, you should just be banned.

Reported for spam and harassment. Have a nice day.

Cry me a river))Resorting to reports, calls for bans, and guessing whether I’m a bot just proves you have absolutely zero technical arguments left. You are acting like a toddler throwing a tantrum because someone took their candy away, completely unable to process how the technology actually works under the hood.You completely dodged every single technical point about attention complexity, interpolation, and visual projection matrices. Your entire meltdown is the ultimate proof that I was right from the start. You're just an empty-headed marketing victim with a massive ego, whining on Hugging Face because you can't distinguish a baseline text architecture from packaged cloud service APIs. Keep coping and hitting that report button.)))

Not a single AI model in the world has a true, honest 1M token context window. Everything beyond that is just a hack relying on interpolation, RoPE scaling, or sparse attention. The same goes for multimodality. You can easily hook up a vision encoder or whatever you want to Qwen yourself. It’s not native, it’s just modular.

For readers of this thread: the claim above that vision can be "simply attached by yourself" is technically invalid — without million-sample datasets and heavy fine-tuning of the projector, a model won't become multimodal. The same goes for 1M context. Training goes 8k -> 32k -> 128k -> 256k -> 1M. FLOPs increase not linearly, but quadratically. Also, the troll devalues the engineering behind a million-token context for no reason at all.

By the way, Kimi K3 was released with their own projector (the advantage of their projector specifically is that it processes images natively without sub-image batching/splitting), they trained it themselves, there's even a paper: https://arxiv.org/html/2504.07491v3, but the one in that paper is used in K2.5. And K3 uses a new version of their projector https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf. Kimi K2.5 was released before Qwen 3.5; even after Qwen 3.5 was released, the vision capabilities of Kimi K2.5 specifically were the best among open-weight models. Yes, previously Qwen used siglip2, but suddenly now they decided to train their own projector too? Maybe that explains why it's in second place on arena.ai in the vision category.

To add further context on the sheer scale of compute required here:
Kimi used 15 Trillion tokens for joint multimodal pre-training and 1 Trillion tokens for MoonViT alone. Look at DeepSeek-V4 — they were the first to pioneer native 1M context in open weights, pre-training on 32 Trillion tokens and designing brand-new attention architectures (CSA & HCA) just to make 1M context computationally viable, followed by MiMo V2.5 and others.
Moonshot similarly spent years experimenting with multimodal fusion, scaling from small prototypes up to trillion-parameter MoE architectures.
Pretending that 1M context and SOTA vision are trivial "plug-and-play hacks that anyone can attach in 5 minutes" completely disregards the massive, multi-million-dollar engineering efforts of the entire open-source community.

Interesting point. Many “million-token” claims do depend on techniques like RoPE scaling, attention optimizations, retrieval, or other engineering trade-offs rather than a model attending equally well to every token. At the same time, those methods are part of modern model design, so the real question is not just context length on paper, but how much useful information the model can reliably retain and reason over across that context. The benchmark should be effective performance, not the advertised token number alone.

Has anyone tried Muse Glimmer 30b? I did and it is far over Qwen 3.6 27b.
I was hopping to see a mini Mamuth in Qwen 3.8 27b.

Muse Glimmer 30b unsloth XL Q8 and bartowski Q8 - both are very very bad if you ask to make "zombie survival" game using tokenchaser promt avilable on his website this LLM fails, cant fix it's own errors etc. Deleted both versions. What a piece of crap. Paid Influencers all over youtube post nonsense about it, how good it is. It's total scam with 131k context only.

I want to clarify that I am still disappointed with how Qwen set expectations during the announcement. As I wrote earlier, while Qwen previously kept the 1M context behind a paywall, the list of restricted features has expanded here—even though 1M context has become the industry norm after DeepSeek V4.

However, looking at the bigger picture: this is only the second model family exceeding 2 trillion parameters, approaching nearly 3 trillion. Architecturally, Qwen is different, and we know they are heavily focused on RL-maxing toward a general-purpose model.

Open weights of this scale are becoming the new norm now that Qwen has joined Kimi in releasing them. Even if stripped, this is the first time Qwen has published a "Max" model at all.

Compared to Kimi—where K3 remains behind a subscription—Qwen is still available in their chat interface for free with generous limits. Granted, the outputs don't contain CoT, but the model is accessible and usable right now at a good speed.

Did Meta announce Spark weights after Kimi and Qwen's releases? Not necessarily. At the end of the day, Qwen could have chosen not to release this model or the upcoming 3.8 27B at all, so we should still appreciate that they continue contributing to the open-source community.

I mostly worry about the team, as they read these discussions. When I opened this thread, I didn't expect comments like "What a pity HF doesn't let us leave bad reviews for models," negative reactions stacked on it, or seeing similar harsh sentiment spilling into threads like #28. Guys, that wasn't the intention behind this discussion.

I hope others share this perspective. Perhaps Qwen saw Kimi's release and wanted to capture attention with what they had ready. But since this is a "Max" model, there might genuinely be unique architectural or technical elements they aren't ready to share freely just yet.

It's hilarious to watch how fast you flipped the switch and turned into an open-source diplomat the moment your marketing illusion shattered. You completely panicked, went to Google, and copy-pasted a wall of text about DeepSeek and Kimi tech reports just to hide the fact that you didn't even know what RoPE or linear attention complexity was five minutes ago.You completely ignored my point because you have zero technical depth. Copy-pasting numbers about 'Trillions of tokens' and 'joint pre-training' doesn't magically create a true linear (O(N^2)) attention matrix for 1M tokens on standard hardware. No matter how many papers you name-drop, a context window scaled via interpolation or sparse block architectures remains a computational workaround, not a 'native 1M linear span' you were crying about.You started this thread throwing a childish tantrum because you didn't get your free '1M context' toy, and now you’re backpedaling, editing your posts, and begging the community to 'appreciate the team' because you realized you look completely clueless. Your total lack of actual architectural understanding combined with this massive ego is pure comedy. Keep editing your comments, it won't change the baseline physics of transformers.

😁 Мне жаль, что ты мы владем с тобой одним и тем же языком, но ты все-таки хотя-бы смешной.

😁 Мне жаль, что ты мы владем с тобой одним и тем же языком, но ты все-таки хотя-бы смешной.

From throwing a tantrum and calling for bans, to copy-pasting technical reports you don't understand, to now typing broken sentences with shaking hands while crying behind a smiley emoji. What a flawless cycle of coping.Next time you decide to whine on Hugging Face about marketing specs, make sure you actually understand the basic physics of transformers first, so you don't have to backpedal and embarrass yourself like this. Have fun editing your posts further. Done with you.

😁 Мне жаль, что ты мы владем с тобой одним и тем же языком, но ты все-таки хотя-бы смешной.

From throwing a tantrum and calling for bans, to copy-pasting technical reports you don't understand, to now typing broken sentences with shaking hands while crying behind a smiley emoji. What a flawless cycle of coping.Next time you decide to whine on Hugging Face about marketing specs, make sure you actually understand the basic physics of transformers first, so you don't have to backpedal and embarrass yourself like this. Have fun editing your posts further. Done with you. russian loser.

Let’s do a clear recap of your comment history across Hugging Face, because watching this pattern unfold is quite revealing.

You go from repository to repository—Meta, Qwen, Mistral, MiniMax, Huawei—copy-pasting the exact same script about "Frankenstein models" and "duct tape". You inspect config.json files, see numbers you don't understand, and form hasty conclusions. But when real developers check your claims, the technical arguments fall apart:

1. The $1/\sqrt{26}$ Constant (Meta Muse-Glimmer):
   In Meta’s Muse-Glimmer repo, you claimed 0.19611613513818404 was an uneducated empirical patch applied after gradients blew up. Actual engineers had to point out that it is literally the standard mathematical variance scaling factor $1/\sqrt{26}$ straight from the official documentation. Your response was a quick backpedal claiming people "missed your deeper point."

2. The Multimodal Misconception:
   In Qwen’s repo, you claimed vision capabilities are "just a 5-minute modular adapter anyone can attach with duct tape." Anyone who has actually trained a Vision-Language model knows that joint multimodal pre-training requires trillions of tokens and massive compute.

3. The Xenophobic Escalation:
   When I pointed out these technical inconsistencies, you immediately abandoned technical discussion and dropped explicit xenophobic slurs ("russian loser"). That alone speaks volumes about your approach to debate.

Looking at your interactions across Hugging Face—from Qwen to MiniMax H3—this pattern resembles the classic online persona from South Park Season 20 (Skankhunt42). You rely on outrage and harsh criticism until someone in the MiniMax thread responds with basic politeness, at which point the hostility instantly vanishes. You push people down over minor gaps in knowledge, yet expect appreciation for it.

You lectured me about RoPE and linear attention as if it’s hidden knowledge. I understand how RoPE, interpolation, and sparse attention work. The core point you dodged is that for real applications, what matters is effective context performance and output quality, not pedantic gatekeeping over architecture labels.

Hiding behind config files doesn't replace actual training experience. I've spent more than enough time on this thread. Keep inspecting JSONs—I'm out.

😁 Мне жаль, что ты мы владем с тобой одним и тем же языком, но ты все-таки хотя-бы смешной.

From throwing a tantrum and calling for bans, to copy-pasting technical reports you don't understand, to now typing broken sentences with shaking hands while crying behind a smiley emoji. What a flawless cycle of coping.Next time you decide to whine on Hugging Face about marketing specs, make sure you actually understand the basic physics of transformers first, so you don't have to backpedal and embarrass yourself like this. Have fun editing your posts further. Done with you. russian loser.

Let’s do a clear recap of your comment history across Hugging Face, because watching this pattern unfold is quite revealing.

You go from repository to repository—Meta, Qwen, Mistral, MiniMax, Huawei—copy-pasting the exact same script about "Frankenstein models" and "duct tape". You inspect config.json files, see numbers you don't understand, and form hasty conclusions. But when real developers check your claims, the technical arguments fall apart:

1. The $1/\sqrt{26}$ Constant (Meta Muse-Glimmer):
   In Meta’s Muse-Glimmer repo, you claimed 0.19611613513818404 was an uneducated empirical patch applied after gradients blew up. Actual engineers had to point out that it is literally the standard mathematical variance scaling factor $1/\sqrt{26}$ straight from the official documentation. Your response was a quick backpedal claiming people "missed your deeper point."

2. The Multimodal Misconception:
   In Qwen’s repo, you claimed vision capabilities are "just a 5-minute modular adapter anyone can attach with duct tape." Anyone who has actually trained a Vision-Language model knows that joint multimodal pre-training requires trillions of tokens and massive compute.

3. The Xenophobic Escalation:
   When I pointed out these technical inconsistencies, you immediately abandoned technical discussion and dropped explicit xenophobic slurs ("russian loser"). That alone speaks volumes about your approach to debate.

---

Looking at your interactions across Hugging Face—from Qwen to MiniMax H3—this pattern resembles the classic online persona from South Park Season 20 (Skankhunt42). You rely on outrage and harsh criticism until someone in the MiniMax thread responds with basic politeness, at which point the hostility instantly vanishes. You push people down over minor gaps in knowledge, yet expect appreciation for it.

You lectured me about RoPE and linear attention as if it’s hidden knowledge. I understand how RoPE, interpolation, and sparse attention work. The core point you dodged is that for real applications, what matters is effective context performance and output quality, not pedantic gatekeeping over architecture labels.

Hiding behind config files doesn't replace actual training experience. I've spent more than enough time on this thread. Keep inspecting JSONs—I'm out.

Nice try shifting the goalposts to my comment history after completely failing to provide a single architectural argument about Qwen. Name-dropping standard documentation formulas like (1/\sqrt{26}) or variance scaling doesn't mask your initial ignorance.You started this thread crying that Qwen 'castrated' the model by stripping 'native' 1M context. Now that you’ve been forced to admit it’s all just RoPE interpolation and compute workarounds, you suddenly claim 'architecture labels don't matter, only performance does'. That is the ultimate definition of backpedaling.The reality remains unchanged: no matter how much compute you throw at joint pre-training or how many project reports you copy-paste, multimodal projection layers remain an engineering abstraction layer, not an intrinsic single-space native system. You fell for marketing flyers, got corrected on basic transformer physics, edited your posts in panic, and now you’re running away. Keep dodging the actual math. Thread over.

What are you malding about? The 3.8 open weight models ARE vision enabled...

What are you malding about? The 3.8 open weight models ARE vision enabled...

Open your eyes and look at the actual config.json inside this repository instead of reading graffiti on the walls. This release is strictly Qwen3MoEForCausalLM — it is a pure, text-only base model with zero multimodal layers or vision heads embedded into its architecture.Any vision capability you are seeing is slapped on externally through a standard modular wrapper via Hugging Face pipelines or separate vision encoders. It is NOT native to this model's body. Crying about screenshots while completely failing to understand the difference between core model architecture and an external script pipeline just proves you have no business being on Hugging Face. Go learn how to inspect weights before spamming threads.

before spamming threads.

If that's true – turn off notifications. But you still need to spew your verbal diarrhea and waste time on it. Why? You're not even following your own logic.

Open your eyes and look at the actual config.json inside this repository instead of reading graffiti on the walls. This release is strictly Qwen3MoEForCausalLM — it is a pure, text-only base model with zero multimodal layers or vision heads embedded into its architecture.Any vision capability you are seeing is slapped on externally through a standard modular wrapper via Hugging Face pipelines or separate vision encoders. It is NOT native to this model's body. Crying about screenshots while completely failing to understand the difference between core model architecture and an external script pipeline just proves you have no business being on Hugging Face. Go learn how to inspect weights before spamming threads.

Hardly a nice thing to say to a fellow LM enthusiast 😕
But anyways, it looks like you are right RE the MAX variant, but the 27b variant does seem to be vision-capable directly from release, unless I am misreading something.

I feel like they did it because they dont want novita / some other inference provider to grab the model the milisecond it releases, and then serve it with full capabilities... ie they would allow more competition.