🚀 Introducing Qwen3-4B-Instruct-2507 & Qwen3-4B-Thinking-2507 — smarter, sharper, and 256K-ready! 🔹 Instruct: Boosted general skills, multilingual coverage, and long-context instruction following. 🔹 Thinking: Advanced reasoning in logic, math, science & code — built for expert-level tasks. Both models are more aligned, more capable, and more context-aware. Huggingface： huggingface.co/Qwen/Qwen3-4B… huggingface.co/Qwen/Qwen3-4B… ModelScope: modelscope.cn/models/Qwen/Qw… modelscope.cn/models/Qwen/Qw…

Aug 6, 2025 · 4:16 PM UTC

143
400
122
3,161
Replying to @Alibaba_Qwen
Qwen knows one thing.. open-source the hell out of eveything 😂
138
Replying to @Alibaba_Qwen
Qwen now seems to be the Opensource leader.
2
54
Replying to @Alibaba_Qwen
I think this 4b model is going to be better than the 20B OSS.
2
1
28
Replying to @Alibaba_Qwen
busy weeks for the Qwen team
11
Replying to @Alibaba_Qwen
Qwen is on 🔥 with open source supremacy
2
Replying to @Alibaba_Qwen
Wow real good job!
2
Replying to @Alibaba_Qwen
Could you compare it with Gemma 4B please? Perhaps it is the biggest competitor in this size range
1
Replying to @Alibaba_Qwen
I didn't even finish downloading the previous one 🥹 thank you!
1
Replying to @Alibaba_Qwen
which subset of GPQA is this?
Replying to @Alibaba_Qwen
looks interesting, but always gotta ask how they stack up in real-world scenarios. benchmarks are one thing, but can they actually tackle tough problems?
Replying to @Alibaba_Qwen
Impressive advancements, yet true intelligence isn't just about scale but the nuanced understanding of context and intent.
Replying to @Alibaba_Qwen
The devil works hard, but the Qwen team works harder
1
1
84
Replying to @Alibaba_Qwen
Please give us Qwen3-8B-Instruct-2507! That's the best size. With 4-bit quantization it is possible to run 8B model on Nvidia GPU with just 8 GB VRAM (typical gamer GPU!) or on a cheap $500 laptop with 16 GB total RAM. And please keep it as free of guardrails as the original!
4
81
Replying to @Alibaba_Qwen
OpenAI: Here is our open-weight model. Qwen:
77
GIF
Replying to @Alibaba_Qwen
Finally people can stop talking about OpenAI OSS
42
Replying to @Alibaba_Qwen
Would topical expert models make sense, like one that specializes in Python and a different one that specializes in sales and psychology? I think this could get you 4b or 7b models that perform like bigger models in their specialization.
4
2
1
17
Replying to @Alibaba_Qwen
Yesterday, Silicon Valley cooked Today, China cooked
15
Replying to @Alibaba_Qwen
Wow, really good benchmarks for a SLM. Will be insanely useful for Agents and Workflows. As always, an amazing job by the @Alibaba_Qwen team. Thanks guys, you rock!
8
Replying to @Alibaba_Qwen
Tendies Of Wisdom
 @TendiesOfWisdom
Aug 6
The new Qwen3-4B-Thinking-2507 model (Q8 quant) on my 8GB laptop GPU is the first small local thinking model to solve my logic game puzzle! Congratulations to @Alibaba_Qwen team!
7
Replying to @Alibaba_Qwen
qwen is dropping everyday 🔥🔥
7
Replying to @Alibaba_Qwen
Fantastic work by the Qwen team! The release of Qwen3-4B-Instruct-2507 and Qwen3-4B-Thinking-2507, with their impressive benchmarks and open-source access, is a game-changer. Excited to see how these models will push the boundaries of AI innovation!
6
Replying to @Alibaba_Qwen
@grok explain this post in Neanderthal language
2
6
Replying to @Alibaba_Qwen
How many models does qwen series 3 has at this point! 2025 has been a monumental year for Alibaba
3
Replying to @Alibaba_Qwen
How to convert this Qwen/Qwen3-4B-Thinking-2507 to .task format to run in Google Edge Gallery? It's high time this option is provided.
3
Replying to @Alibaba_Qwen
Nice work!
3
Replying to @Alibaba_Qwen
😮 you’re unstoppable 👏👏
3
Replying to @Alibaba_Qwen
@UnslothAI waiting for your fine-tuning notebooks. This is going to be 🔥🔥🔥
3
Replying to @Alibaba_Qwen
Is it available in qwen chat?
3
Replying to @Alibaba_Qwen
How are you doing this qwen? Can you apply this to the new openai model? 😭
3
Replying to @Alibaba_Qwen
thanks bros
2
Replying to @Alibaba_Qwen
Expert systems have always reasoned better than neural nets, and much more cheaply. No reason not to use generative LLM as a front end to an expert system. Best of both worlds.
1
2
Replying to @Alibaba_Qwen
2
GIF
Replying to @Alibaba_Qwen
demo?
2
Replying to @Alibaba_Qwen
Really hope they drop the 8B model next.
1
Replying to @Alibaba_Qwen
Would love a Coder variant
1
Replying to @Alibaba_Qwen
Curious, why? Where is a 4b param being used ? What is the use case in general? Robot and edge devices can run larger models right? Even an Rk1 can run larger. Just want to know.
3
1
Load more