4 ms·

Qwen3.8-2.4T

84 points by mmastrac 2mo ago

brrrrrm 2mo ago
this is Qwen3.8 max, right? https://qwen.ai/blog?id=qwen3.8 https://qwen.ai/blog?id=qwen3.8
- nateb2022 2mo ago
  That's my bet, I doubt they'd have the resources to have trained anything substantially larger
- efficax 2mo ago
  they explicitly say this is qwen3.8 max in the hugging face page
UncleOxidant 2mo ago
We're all really waiting for the 3.8-27B which is due out in 2 days.
- nateb2022 2mo ago
  Yup. I posted about the huggingface countdown: https://news.ycombinator.com/item?id=49272534 https://news.ycombinator.com/item?id=49272534 but seems like both the huggingface page, and the modelscope.cn URL that commenters found, were quickly taken down. Not sure if the submission being flagged is also the doing of whoever took down the huggingface/modelscope.cn pages. Edit: for anyone curious, I saw the huggingface countdown page for Kimi K3, which gave me a hunch that led me to the countdown page for Qwen3.8-27B. As of ~11am EST today, the countdown page said 2 days and 2 hours to release.
  - spwa4 2mo ago
    Back up: https://modelscope.cn/models/Qwen/Qwen3.8-27B https://modelscope.cn/models/Qwen/Qwen3.8-27B counting down to: 2026-08-14 17:00 CEST 2026-08-14 15:00 UTC 2026-08-14 08:00 PDT
WalterGR 2mo ago
Submitted 3 hours ago with 63 comments so far: https://news.ycombinator.com/item?id=49273478 https://news.ycombinator.com/item?id=49273478
xlayn 2mo ago
I wonder besides big labs and REALLY big corporations who can run the 2.6TB q8... I mean, the 1 bit one is 508GB. Assuming 256k context size as irrelevant at those sizes, I would need 22 AMD 7900XTX (24GB vram) to run this for the 1 bit one, and 113 for the 2.6TB 113 and assumming close to constant load and the gpus taking turns as it does on my machine with two gpus is around 113gpus*113W = 11KW... just to hold and run that one instance... and only god knows the cooling requirements... I guess just someone on Meta/Googly/Claudy will say... oh nice, let's download it and run it... How did the unsloth guy did to process a file that size? https://huggingface.co/unsloth/Qwen3.8-2.4T-A95B-GGUF https://huggingface.co/unsloth/Qwen3.8-2.4T-A95B-GGUF
- lytedev 2mo ago
  But nobody that can run this is running it on that hardware, surely?
- tempest_ 2mo ago
  Newegg lists h200s ~40k another 50k for the machine to put your 8 h200s in Its pricey obviously but the 500-1M price tag is not out of the realm for most large companies.
- YetAnotherNick 2mo ago
  > Meta/Googly/Claudy will say... oh nice, let's download it and run it... You can run it for like <$50/hr. You don't need to be Meta/Googly/Claudy for that. For batch inference it makes sense to run it yourself(high latency, high throughput, predefined workload for few hours or day).
edhan235 2mo ago
I have been testing this for hark.news

Qwen3.8-2.4T

84 points by mmastrac 2mo ago

brrrrrm 2mo ago

this is Qwen3.8 max, right? https://qwen.ai/blog?id=qwen3.8 https://qwen.ai/blog?id=qwen3.8

nateb2022 2mo ago
That's my bet, I doubt they'd have the resources to have trained anything substantially larger
efficax 2mo ago
they explicitly say this is qwen3.8 max in the hugging face page

UncleOxidant 2mo ago

We're all really waiting for the 3.8-27B which is due out in 2 days.

nateb2022 2mo ago
Yup. I posted about the huggingface countdown: https://news.ycombinator.com/item?id=49272534 https://news.ycombinator.com/item?id=49272534 but seems like both the huggingface page, and the modelscope.cn URL that commenters found, were quickly taken down. Not sure if the submission being flagged is also the doing of whoever took down the huggingface/modelscope.cn pages. Edit: for anyone curious, I saw the huggingface countdown page for Kimi K3, which gave me a hunch that led me to the countdown page for Qwen3.8-27B. As of ~11am EST today, the countdown page said 2 days and 2 hours to release.
- spwa4 2mo ago
  Back up: https://modelscope.cn/models/Qwen/Qwen3.8-27B https://modelscope.cn/models/Qwen/Qwen3.8-27B counting down to: 2026-08-14 17:00 CEST 2026-08-14 15:00 UTC 2026-08-14 08:00 PDT

WalterGR 2mo ago

Submitted 3 hours ago with 63 comments so far: https://news.ycombinator.com/item?id=49273478 https://news.ycombinator.com/item?id=49273478

xlayn 2mo ago

I wonder besides big labs and REALLY big corporations who can run the 2.6TB q8... I mean, the 1 bit one is 508GB. Assuming 256k context size as irrelevant at those sizes, I would need 22 AMD 7900XTX (24GB vram) to run this for the 1 bit one, and 113 for the 2.6TB 113 and assumming close to constant load and the gpus taking turns as it does on my machine with two gpus is around 113gpus*113W = 11KW... just to hold and run that one instance... and only god knows the cooling requirements... I guess just someone on Meta/Googly/Claudy will say... oh nice, let's download it and run it... How did the unsloth guy did to process a file that size? https://huggingface.co/unsloth/Qwen3.8-2.4T-A95B-GGUF https://huggingface.co/unsloth/Qwen3.8-2.4T-A95B-GGUF

lytedev 2mo ago
But nobody that can run this is running it on that hardware, surely?
tempest_ 2mo ago
Newegg lists h200s ~40k another 50k for the machine to put your 8 h200s in Its pricey obviously but the 500-1M price tag is not out of the realm for most large companies.
YetAnotherNick 2mo ago
> Meta/Googly/Claudy will say... oh nice, let's download it and run it... You can run it for like <$50/hr. You don't need to be Meta/Googly/Claudy for that. For batch inference it makes sense to run it yourself(high latency, high throughput, predefined workload for few hours or day).

edhan235 2mo ago

I have been testing this for hark.news