Qwen3 will be released in the second week of April

[-]

silenceimpaired@reddit

Hi, I’m from the future… the release was both exciting and frustrating. Part of the frustration was around licensing, and the other part was around model sizes.

[-]

I've read a paper stating that sota's model "time horizon" (the time for a human to achieve a task that a sota's model achieve at 50% rate) double every 7 months. Last qwen release was 7 months ago? Lol.

Btw has someone noted that source? I lost it. There was a post here like 3 days ago

[-]

social_tech_10@reddit

https://arxiv.org/abs/2503.14499

[-]

No_Afternoon_4260@reddit

Exactly! Thx

[-]

brown2green@reddit

Was it this one? https://arxiv.org/abs/2412.04315v1

[-]

tengo_harambe@reddit

Can't believe Qwen2.5 was released only 6 months ago. Feels like it's been years, what a journey it's been through. High hopes that Qwen3 takes up the mantle for the next generation of open source.

[-]

TheTerrasque@reddit

The original ChatGPT was released November 30, 2022 - about 2.5 years ago. Feels like 10 years

[-]

Healthy-Nebula-3603@reddit

Yes and I thought the quality like had GPT 3.5 then we get something similar in quality in 5 years on home pc .... oh boy I was sooo wrong

[-]

TheTerrasque@reddit

imagine what it will be in 10 years

[-]

Healthy-Nebula-3603@reddit

With present development AI .... we can't predict 2 years ahead and you're telling about 10?

[-]

pkmxtw@reddit

Qwen2.5 is still pretty much SOTA in every size category.

[-]

Healthy-Nebula-3603@reddit

ekm ... QwQ 32b ...

[-]

JockY@reddit

Yup, 72B @ 8bpw is still my daily driver.

[-]

mpasila@reddit

I'm hoping it'll be as multilingual as Gemma 3 is.

[-]

frankh07@reddit

Will there be a significant breakthrough? It wasn't long ago that Qwen 2.5 was released.

[-]

pseudonerv@reddit

haha, Meta-llama-4 will never see the light of day...

[-]

MorallyDeplorable@reddit

still waiting on sonnet 4 and gpt 5 too

[-]

CarbonTail@reddit

Still remember him asking for $7 TRILLION to invest in compute hardware, lmfao.

[-]

bblankuser@reddit

sam said months for 5

[-]

mosthumbleuserever@reddit

GPT-5 was announced for May. People keep asking where it is and it's been on schedule so far this whole time 🙄

[-]

dark-light92@reddit

At this rate Deepseek V5 & Llama 5 is going to come before GPT 5.

[-]

SelectionCalm70@reddit

but openAI is gonna open source one of their model to compete in open source world

[-]

webitube@reddit

Well, now, uh, Greg, Bret, and I wait until nightfall, and then leap out of the rabbit, taking the community by surprise -- not only by surprise, but totally unarmed!

[-]

AryanEmbered@reddit

Are you sama?

[-]

Silver-Champion-4846@reddit

is that a hint about making llms natively optimized to run on arm cpus?

[-]

-p-e-w-@reddit

It’s crazy when you realize what is ultimately happening here. Meta, with near-unlimited funds and tens of thousands of elite engineers, can’t compete at the top level anymore. The top Chinese players have pushed out multiple amazing models each since Llama 3.1.

[-]

pseudonerv@reddit

Exactly. Closed models and national boundaries fragment innovation and slow down progress. Imagine how far we could go if brilliant minds worldwide collaborated openly instead of working in isolation.

[-]

TheRealGentlefox@reddit

Meta's last model release was easily SotA for its size class.

[-]

noage@reddit

Compute is still very important and applying lessons from all those models to a base with as much compute put into it as meta can muster could be amazing. The bitter lesson is often posted here which speculates that human ingeniity to work on ai type can benefit but it's the compute that ultimately proves more successful time and again.

[-]

SelectionCalm70@reddit

sed to see llama which started as a torch bearer for open source world is nowhere in the race

[-]

RMCPhoto@reddit

Nowhere in the race?

Llama 3.3 was just released in December and is at the same level as qwen 2.5 and is still sota for ifeval.

Llama 3.2 is in a similar boat.

[-]

TheRealGentlefox@reddit

I honestly wonder if there is some astroturfing or something in this sub that pushes "Meta sucks, China #1" ideas. (Not saying OP is part of that) Hell, we're still getting people reveling in the idea of the Meta team seething at all these other releases.

Like not only did Meta create the open LLM scene (hence our sub name) and hasn't shafted us yet, but their model drops have always been amazing. I had to point out just yesterday that 405B beats or ties the initial version of 4o, Mistral Large, and Claude 3 Opus. Llama 3.2 was the only model I've been able to run on my phone that didn't shit its pants in stupidity until Gemma 3.

I love that China is doing open-weight, Wenfeng is my favorite of the LLM CEOs, and DeepSeek is absolutely killing it right now, but let's not demonize all the Western teams and pretend Chinese ones are leaving them in the dust.

[-]

Current1034785@reddit

No. It may be close in some benchmarks but Llama is not in the same league for coding and other major use cases in practical use. Llama's outputs are short and simple in comparison.

[-]

Dry-Judgment4242@reddit

Llama 3.3 listen to context better like a good boi. But Qwen2.5 more unruly and street smart.

[-]

usernameplshere@reddit

Hopefully we will get the final release of QwQ Max then as well.

[-]

Healthy-Nebula-3603@reddit

you mean next version ;)

[-]

secopsml@reddit

my GPU asks for new coolant

[-]

Enough-Meringue4745@reddit

Need more vespene gas

[-]

throwawayacc201711@reddit

Construct additional pylons

[-]

Cute_Translator_5787@reddit

Not enough minerals

[-]

Substantial-Ebb-584@reddit

Then localLLM: Spawning more overlords

[-]

ThinkExtension2328@reddit

Wait hang on a min, could actual coolant be used for a computer ? Why don’t we?

[-]

_supert_@reddit

There was a trend in bitcoin mining asics to cool by submerging in a liquid.

[-]

Mice_With_Rice@reddit

It's used all the time. Heat pipes use a gas as the coolant. Water loops use you know what as the coolant.

[-]

lack_of_reserves@reddit

Wait. Water loops use gasoline as coolant?

[-]

BlackmailedWhiteMale@reddit

Gas coolant, steam as a gas.

[-]

Mice_With_Rice@reddit

Pretty close. The copper pipes use steam. Aluminum pipes use ammonia. (typically). There are different mediums possible for them.

[-]

Mice_With_Rice@reddit

🙄 The verbiage is different depending on what country you live in. But water means water in every common vernacular.

[-]

dergachoff@reddit

— Uses too many big words

[-]

No_Kick7086@reddit

Awesome to see this

[-]

Ok_Landscape_6819@reddit

Nice Llama 4 and Qwen3 this month. Also R2 maybe ? And GPT-5 next month ; Next two months will be wild..

[-]

Current1034785@reddit

Do we have any clues if open weights and parameter sizes?

[-]

xqoe@reddit

When 15B passive 2B active

[-]

keepthepace@reddit

Hot take: We should downvote announcements to minimize hypes. Talk about releases, not announcements.

[-]

vibjelo@reddit

Also, considering we don't even know if they'll actually release any weights so you can run it locally, it might belong here even less.

[-]

Such_Advantage_6949@reddit

Hope this wont delay llama4 further

[-]

vibjelo@reddit

If Llama is being delayed because others keep releasing actually open source weights that are better than Llama, then I hope it keeps getting delayed forever. Rather have high quality open models than whatever Meta keep trying to push.

[-]

Sambojin1@reddit

I'm hoping they do a little 5B model for edge devices. Better than 3B, but faster than 7-8-9B, yet still fits on anything.

[-]

Longjumping-Solid563@reddit

Here's a little bit of what we know

https://huggingface.co/Qwen/Qwen3-15B-A2B (MOE model)

https://huggingface.co/Qwen/Qwen3-8B-beta

Qwen/Qwen3-0.6B-Base

Qwen3-15B-A2B is very promising. Most phones can load 15b quants and with 2b active params it will perform well. It would be fucking sick if it can run practically on a 16gb Pi 5. But I don't think we've seen a successful MOE model at this size (Unless closed lab), so grain of salt of course. Have you tested out the LG models at all, they look promising for edge too.

[-]

Gold_Ad_2201@reddit

LG models give terrible answers for me. not even close to qwen

[-]

singinst@reddit

Really? LG models are pretty solid IMO... similar to Qwen2.5... but QwQ prob better still

[-]

Gold_Ad_2201@reddit

maybe it is limited to my workloads - coding tasks. exaone got into repeat loops a lot and I just gave up and went back to qwen

[-]

ReMeDyIII@reddit

Isn't Qwen usually censorship heavy?

[-]

My_Unbiased_Opinion@reddit

Depends. Politically yes, but practically, not really. It doesn't have much issues giving financial advice or even medical information.

[-]

Acrobatic_Cat_3448@reddit

'most significant model product in the first half of 2025'? So what's to happen in the second?

[-]

shroddy@reddit

Conservative estimate: Qwen3-VL and Qwen3.5

Realistic estimate: a multimodal model with image Gen capabilities

Really hopeful estimate: which is as good as current chat gpt.

[-]

TheSilverSmith47@reddit

Very evil of you to post this on April 1st

[-]

umarmnaq@reddit

It's second

[-]

Orolol@reddit

It's already the third here.

[-]

RamboLorikeet@reddit

[-]

umarmnaq@reddit

huh?

[-]

Mice_With_Rice@reddit

"Tiger Sniff"... is that a real name or wonky machine translation?

[-]

CLST_324@reddit

Wonky machine translation. Though 虎嗅 has its special meaning, it's just better to call it Huxiu in English.

[-]

Glad-Cook-3539@reddit

Weird and useless knowledge:

The phrase "In me the tiger sniffs the rose" from a poem by Siegfried Sassoon was translated into Chinese by Yu Kwang-chung as "心有猛虎细嗅蔷薇" (literally: "heart has fierce tiger, gently sniffing roses"). the translation has become a widely circulated poetic expression in the Chinese-speaking world.

[-]

AaronFeng47@reddit (OP)

"Tiger Sniff" is technically correct, just sounds weird in English

[-]