Qwen Blog

Outside of Google Deepmind open sourcing the code and weights of AlphaFold, I don't think they've released any of their GenAI stuff (Imagen, Gemini, Flash 2.5, etc). The best multimodal models that you can run locally right now are probably Qwen-Edit 20b, and Kontext.Dev. https://qwenlm.github.io/blog/qwen-image-edit https://bfl.ai/blog/flux-1-kontext-dev

The PR confuse a but 32k/64k and 131k if paid API. Also this model https://huggingface.co/Qwen/Qwen3-235B-A22B Is native 32k. So the 64k and 131k use ROPE that is not the best for effective context. While https://qwenlm.github.io/blog/qwen3-coder/ it's 256k native https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct.

Lots of folks working on open-source reasoning models trained with reinforcement learning right now. The best one atm appears to be Alibaba's 32B-parameter QwQ: https://qwenlm.github.io/blog/qwq-32b-preview/ I also recently wrote a blog explaining how reinforcement fine-tuning works, which is likely at least part of the pipeline used to train o1: https://openpipe.ai/blog/openai-rft

Couple notes for newcomers: 1. This is a VLM, not a text-to-image model. You can give it images, and it can understand them. It doesn't generate images back. 2. It seems like Pixtral 12B benchmarks significantly below Qwen2-VL-7B [1], so if you want the best local model for understanding images, probably use Qwen2. If you want a large open-source model, Qwen2-VL-72B is most likely the best option. 1: https://qwenlm.github.io/blog/qwen2-vl/

I thought it'd be interesting to test qwen3 uncensored given the fantastic benchmarks (https://qwenlm.github.io/blog/qwen3/) and realized the barrier to entry for some folks may be high. I've improved qwen3_8B (named Swift) and qwen3_32B (named Eagle) for hassle-free responses and made access as easy as possible for hacker news folks. The quick responses from Swift are satisfying. These models are run locally on my machine with burned logs. No data is saved or passed on the server side. Hit the nuke button from the menu to 100% clear the client side. If my GPUs start melting, I'll have to password protect the site. I hope some folks can explore their thoughts freely before then. Open to feedback. The free link is just for yc/hackernews folks https://speaksy.chat/yc Please use responsibly. There are so many curious thoughts we have that get blocked from exploring via flagship models and I understand why. e.g. in the security space. However, I think curious minds demand an outlet. If you see anything awful, please use the Report Abuse feature by clicking the triple dot on any message.

Qwen2-VL is the latest version of our vision language models built upon Qwen2. It consists of the following features: SoTA understanding of images of various resolution & ratio: Qwen2-VL achieves state-of-the-art performance on visual understanding benchmarks, including MathVista, DocVQA, RealWorldQA, MTVQA, etc. Understanding videos of 20min+: Qwen2-VL can understand videos over 20 minutes for high-quality video-based question answering, dialog, content creation, etc. Agent that can operate your mobiles, robots, etc.: with the abilities of complex reasoning and decision making, Qwen2-VL can be integrated with devices like mobile phones, robots, etc., for automatic operation based on visual environment and text instructions. Multilingual Support: to serve global users, besides English and Chinese, Qwen2-VL now supports the understanding of texts in different languages inside images, including most European languages, Japanese, Korean, Arabic, Vietnamese, etc. dang please update URL https://qwenlm.github.io/blog/qwen2-vl/ More details > A key architectural improvement in Qwen2-VL is the implementation of Naive Dynamic Resolution support. Unlike its predecessor, Qwen2-VL can handle arbitrary image resolutions, mapping them into a dynamic number of visual tokens, thereby ensuring consistency between the model input and the inherent information in images. This approach more closely mimics human visual perception, allowing the model to process images of any clarity or size

Qwen3 was trained for tool usage too. Most models are these days. https://qwenlm.github.io/blog/qwen3/#agentic-usages

https://qwenlm.github.io/blog/qwen3/ Go look at the benchmark numbers of qwen3-4B if you think these are unrealistic.

Just add "/no_think" in your prompt. https://qwenlm.github.io/blog/qwen3/#advanced-usages

I'm not sure to which extent this opinion is accurately informed. It is well known that nobody trains on 1M token-long content. It wouldn't work anyway as the dependencies are too far fetched and you end up with vanishing gradients. RoPE (Rotary Positional Embeddings, think modulo or periodic arithmetics) scaling is key, whereby the model is trained on 16k tokens long content, and then scaled up to 100k+ [0]. Qwen 1M (who has near perfect recall over the complete window [1]) and Llama 4 10M pushed the limits of this technique, with Qwen reliably training with a much higher RoPE base, and Llama 4 coming up with iRoPE which claims scaling to extremely long contexts up to infinity. [0]: https://arxiv.org/html/2310.05209v2 [1]: https://qwenlm.github.io/blog/qwen2.5-turbo/#passkey-retrieval

I think Qwen team is alibaba AI arm https://qwenlm.github.io/about/

See QwQ-Max-Preview: https://qwenlm.github.io/blog/qwq-max-preview/

Wasn't this release in Nov 2024 as a "preview" with similarly impressive performance? https://qwenlm.github.io/blog/qwq-32b-preview/

I'd really like to play with Qwen2.5-VL at some point, perhaps for reading data-sheets for microchips. Nicely for some applications, it's also very good at reporting position of what it finds, which many ML tools are pretty mediocre at. https://qwenlm.github.io/blog/qwen2.5-vl/ Not really this application, but QvQ for visual reasoning is also impressive. https://qwenlm.github.io/blog/qvq-72b-preview/ Meta has used Qwen as the basis for their Apollo research. https://arxiv.org/abs/2412.10360

> Unfortunately Gemini really seems to struggle on this, and no matter how we tried prompting it, it would generate wildly inaccurate bounding boxes Qwen2.5 VL was trained on a special HTML format for doing OCR with bounding boxes. [1] The resulting boxes aren't quite as accurate as something like Textract/Surya, but I've found they're much more accurate than Gemini or any other LLM. [1] https://qwenlm.github.io/blog/qwen2.5-vl/

Actual announcement: https://qwenlm.github.io/blog/qwen2.5-max/

Maybe link to the original announcement instead of an Apidog advertisement: https://qwenlm.github.io/blog/qwen2.5-max/

Blog: https://qwenlm.github.io/blog/qwen2.5-max/ Seems like the model is not available for download on HF/Github

This API only model with a 1M context window was released back in Nov. Just for some historical context. https://qwenlm.github.io/blog/qwen2.5-turbo/

The official announcement from the Qwen team is also on HN right now. https://news.ycombinator.com/item?id=42831769 https://qwenlm.github.io/blog/qwen2.5-1m/

Non-blogspam link: https://qwenlm.github.io/blog/qwq-32b-preview/ discussion: https://news.ycombinator.com/item?id=42259184

> The issue with some recent models is that they're basically overfitting on public evals… You really want to be testing stuff that isn't overfit to death… I'd be shocked if this model held up in the comprehensive private evals. From the announcement: > we selected the latest 4 months of LiveCodeBench (2024.07 - 2024.11) questions as the evaluation, which are the latest published questions that could not have leaked into the training set, reflecting the model’s OOD capabilities. — https://qwenlm.github.io/blog/qwen2.5-coder-family/

I'd also like to point out that they omit Qwen2.5 14B from the benchmark because it doesn't fit their narrative(MMLU Pro score of 63.7[0]). This kind of listing-only-models-you-beat feels extremely shady to me. [0] https://qwenlm.github.io/blog/qwen2.5/

This might be what you are asking for... https://qwenlm.github.io/blog/qwen2.5-coder/ Ctrl F - Code Reasoning:

Qwen MoE[0] showed promise for a small size. I hope they spend more time in them. [0] https://qwenlm.github.io/blog/qwen-moe/

They are actually very good, cf: https://qwenlm.github.io/blog/qwen-vl/ and it's possible to unalign them (like Dolphin-Yi).

Blog post: http://qwenlm.github.io/blog/qwen-image/

I've been following the state of things, but I'm not sure which ones are the best. There's Meta's CodeLlama[1], Mistral's Codestral[2], DeepSeek AI's DeepSeek-Coder-V2-Instruct[3], CodeGemma[4], Alibaba's CodeQwen[5], and Microsoft's WizardCoder[6]. I'm pretty sure CodeLlama is out of date now. I've heard DeepSeek LLMs are good and DeepSeek-Coder-V2-Instruct was released recently. With the good reputation and its massive size (236b) I'd guess it is the best coding LLM, but if it's not being trained efficiently, maybe Codestral and Codestral Mamba come close. I don't think the best coding LLMs are close to GitHub Copilot but I could be wrong since I'm just relaying information that I've heard secondhand. [1] https://ai.meta.com/blog/code-llama-large-language-model-coding/ [2] https://mistral.ai/news/codestral/ [3] https://github.com/deepseek-ai/DeepSeek-Coder-V2 [4] https://developers.googleblog.com/en/gemma-family-expands-with-models-… [5] https://qwenlm.github.io/blog/codeqwen1.5/ [6] https://github.com/nlpxucan/WizardLM

See also: - Paper page: https://huggingface.co/papers/2512.15603 - Model page: https://huggingface.co/Qwen/Qwen-Image-Layered - Quantized model page: https://huggingface.co/QuantStack/Qwen-Image-Layered-GGUF - Blog URL: https://qwenlm.github.io/blog/qwen-image-layered/ (404 at the time of writing this comment, but it'll probably release soon) - GitHub page: https://github.com/QwenLM/Qwen-Image-Layered