You must log in or register to comment.
In my informal testing, Qwen72b is quite good. I anecdotally rate it stronger than Llama 2 from the few tests that I have conducted.
What tests have you tested this in?
I’m very interested in storytelling and RP
can it beats 3.5-turbo?
Would it be possible to merge it with deepseek coder 35B?
Is it censored?
This is amazing. Yesterday we got Deepseek, and today we’re getting Qwen. Thank you for releasing this model!
I’m looking forward to seeing comparisons
Is there any free website where I can test those Chinese models? Thanks.
for Deepseek there is https://chat.deepseek.com/ , for qwen I don’t know.
The first thing I looked for was the number of training tokens. I think yi34 got a lot of benefit from 3 trillion, so this model having 3 trillion bodes well.
Their own quants for transformers run:
https://huggingface.co/Qwen/Qwen-72B-Chat-Int8
https://huggingface.co/Qwen/Qwen-72B-Chat-Int4https://github.com/QwenLM/Qwen
Also released was a 1.8B model.
From Bunyan Hui’s Twitter announcement:
“We are proud to present our sincere open-source works: Qwen-72B and Qwen-1.8B! Including Base, Chat and Quantized versions!
🌟 Qwen-72B has been trained on high-quality data consisting of 3T tokens, boasting a larger parameter scale and more training data to achieve a comprehensive performance upgrade. Additionally, we have expanded the context window length to 32K and enhanced the system prompt capability, allowing users to customize their own AI assistant with just a single prompt.
🎁 Qwen-1.8B is our additional gift to the research community, striking a balance between maintaining essential functionalities and maximizing efficiency, generating 2K-length text content with just 3GB of GPU memory.
We are committed to continuing our dedication to the open-source community and thank you all for your enjoyment and support! 🚀 Finally, Happy 1st birthday ChatGPT. 🎂 “
we have expanded the context window length to 32K
Kinda buried the lead here. This is far and away the biggest feature of this model. Here’s hoping it’s actually decent as well!
Well, it depends on how well it keeps the context resolution. Did you see that comparison sheet on Claude and GPT-4? Astounding.
my heart skipped a beat because I thought it was Qwen-1.8T.
If the US keeps going full woke and are too afraid to work as hard as possible on the LLM ecosystem, China won’t wait twice before winning this battle (which is basically the 21th century battle in terms of technology)
Feels sad to see the US decline like that…
It would be great to see gguf versions. (At least, my workflow right now goes via ollama.) How are people running Qwen-72B locally right now?
Amazing capabilities on some benchmarks if true.
big if true
Bit disappointed by the coding performance but it is a general use case model. It’s insane how good gpt 3.5 is for how fast it is.
Apparently the chat version has about 64 for humaneval.
What do these tests mean for LLM? There are many values, and I see that in most cases qwen is better than gpt4. In others it is worse or much worse
All the cases it is better than GPT-4 are benchmarks involving Chinese language. OpenAI is going to have a hard time getting access to extensive Chinese language datasets so it’s not surprising a 72B model can beat GPT-4, though it’s still impressive in it’s own right.
Now everyone is most interested in how much better it is than 70b llama
Very nice.
I wonder what the performance degradation is after quantising. For other models some users reported that quantizing greatly affected other language capabilities and this model seems to be at leash 50% Chinese.
I’ve seen that ChatGLM began talking in mixed Chinese/English when asked “What tips do you have for a mountaineering trip?”
Just when I’d talked myself out of getting a second 3090
