Qwen-2 : The BEST Opensource LLM is here & It's Amazing! (Beats Llama-3, GPT-4O, Claude)
HTML-код
- Опубликовано: 6 июн 2024
- In this video, We'll be talking about the newly released Qwen-2 models. The models include 0.5B Parameter model, 1.5B Parameter Model, 7B Parameter model and upto 72B Parameter model. The first three models can even run on computers with out any need of GPUs which is amazing. You can make very simple and easy to use Copilot with it as well. It beats every other model such as Llama 3, Claude-3, Gemini & others in multiple benchmarks. You can use it as a Github Copilot alternative with your VSCode or Neovim setup.
[Resources]
Qwen-2 Release Blog Post: qwenlm.github.io/blog/qwen2/
[Key Takeaways]
📢 Qwen-2 Released: Alibaba's Qwen-2 models are now publicly available, bringing new advancements in AI and machine learning.
📊 Multiple Model Variants: Qwen-2 offers various sizes, including 0.5B, 1.5B, 7B, 57B, and 72B parameters, catering to different needs and capabilities.
🌐 Multilingual Support: These models support 27 languages in addition to English and Chinese, making them versatile for global use.
🚀 Extended Context Length: Qwen-2 models now support up to 128K tokens, enhancing their ability to handle complex tasks and large datasets.
⚡ Enhanced Performance: Claiming state-of-the-art results, Qwen-2 excels in benchmarks for coding, mathematics, and more, outperforming many competitors.
💻 Open-Source Availability: Qwen-2 models are available on HuggingFace, Modelscope, and Ollama, making them accessible for developers and researchers.
🔍 Smaller Models, Big Impact: The smaller models (0.5B, 1.5B) show impressive capabilities, promising efficient performance on everyday devices. - Наука
Really great content, thank you! Just discovered your channel and really loving it - despite your narration sounding like you're not going to open the pod bay doors for me, the detail and meat are just fantastic, just subscribed 🙏💪
It's an AI voice
@@altdoom5205 I figured - and it's a great choice, just really reminds me of HAL - but in a good way :)
Thanks great video bro. This news are always welcome!
Great videos. Keep going.
Nice video! you could put some link into the description, like the benchmark test, huggingface and your own video about vscode.
How is this beating gpt-4o?
Thank you so much for valuable information
copilot vscode link?
Really good video. I saw your video via my curiosity to learn more about this open-source models. I'm a programmer but a noob in AL&ML topics.
I want to run my own local LLM but I also want it to fetch latest information from the internet in case I ask it a questions which requires up-to-date information.
I believe you explained in one of your videos that it's called a RAG model where the LLM first contacts its source-of-information (internet) and then combines it with its own prompt.
If you can make some videos explaining how we can achieve that in our local computer, that would be amazing. Also waiting for you co-pilot video with qwen-2.
Thank you for what you do. I hope you have a nice day.
I'm trying to break into this space but one thing I can't put a bounds on is the difficulty to make and use these kinds of models in normal CPU specs...this one can? How come everyone isn't doing this?
I think with 3B or less models it is feasible, the performance will not be extreme tho.
Yes. I want copilot with these models please :D. It's really amazing if 0.5B parameter can also do it. My expectation were low though... But still hoping lol.
I gave it a quick test and it seems broken to me, but maybe i have to test it better.
Can anyone confirm?
Yes can you try and install it locally and make it your own copilot and test it against the real copilot in a let's say any large open source repo and since this model is basically the best at 7b params can you also show us how to apply Rag for this one
Please show a copilot video with this models. The small ones more precisely.
Thank you!
How does it compare with codemistral?
Codestral is better for coding than Qwen 70b. I tested on Edabit on expert challenges.
I have tryed the 7B model in LM studio, and it only writes nosense ...
Exactly! Spitting nonsense
What preset are you using?
@@mackroscopik several, one of them the LM Studio default
Yeah bro. Its superb. Continue this with copilot...
I tested it as well, and I don‘t think that it beats GPT 3.5. Yes, It‘s quick, but its German is not that good. Lots of lexic and grammar mistakes. I have to test it more on reasoning, because it could be the ultimate substitute for llama3:8b if I want to analyze a prompt. (Q&A: Yes, certain phrase + no)
It's not meant to be multilingual it's trained on English and Chinese. There is some other languages in the dataset, but that wasn't their main focus for this model.
@@figs3284 , excuses. Got you paid for that?
我对中文还比较满意
@@3a146 , I’m glad for you. The model is btw great for inline-decisions.
@@MeinDeutschkurs Haha, seems babel tower is already here!
So thrill hearong qwen2 beat llama3... cant wait to try it
run through terminal and enchanted using ollama. D**n, its so fast, and that haystack is no jokes. Now I'm tryn to build full webapp... lets see how its goin
What's the configuration of your computer
Potato CPU with a non existing GPU inside an IBM case. I'm happy with it.
In numbers
Is Qwen2 censored or uncensored? This should be the first thing mentioned about a model, because if it's censored, it's useless :)
That voice 😅🪬 subscribed 😅✌
its on LM STUDIO, probably the fastest 7B model I have installed on LM STUDIO.
Yes, it's one of the fastest model. I think it's because of GQA.
"small models are the future", not until they are as accurate and don't hallucinate so damn much LLAL
Hallucinations are a feature, not a bug. Use critical loop + external tools to fact check and correct hallucinations, use hallucinations as random seeding for crativity.
is this better than Codestral 22B ?
According to my tests, no
@@MudroZvon Have you tried the Qwen 2 70B Cognitive version ? Apparently its fully uncensored and might even be smarter.
@@MudroZvon how do you test these models by the way?
@@aneeshprasobhan I tested Cognitive version on "Huggingface" too. Didn't notice any improvement in coding. I tested them on several "Edabit" expert tasks. Codestral on Mistral's "Le chat" is significantly better. I even managed to solve a couple of expert tasks in one shot. GPT-4o is able to do that as well.
So the benchmarks for Qwen-72b are overstated, in my opinion. I wouldn't say it's better than LLAMA 3 either. It fails simple reasoning questions like "Today I have 32 apples. Yesterday I ate 28. How many apples do I have?". But it's still a pretty interesting model. It's very cheap on Poe AI (15 points for request from daily 3000 points)
The qwen2 72B is very fast with 2 GPUs...
yes please....
❤❤❤❤❤
You are the brother of GLaDOS from Portal 2 lol
awesome ! :D
Beats gpt4o? C'mon man
But really. In Layla lite, qwen 2 1.5b can run in my phone and much smarter than tinyllama.
Sadly it can't answer controversial topics...
NousResearch/Hermes-2-Theta-Llama-3-8B is better, Queen Failed this question for me:
There is a cake on a table in the dininig room, I walk over to the cake and place a plate on top of the cake, I then pick up the plate and take it into the kitchen.
Which room is the cake currently in?
Hermes-2 gets it right every time