Summary: The big AI events of November

Posted by nh_local@reddit | LocalLLaMA | View on Reddit | 21 comments

* **Alibaba** released its new model, **QwQ 32B Preview**, which integrates reasoning capabilities before responding. The model competes with, and sometimes surpasses, OpenAI's **o1-preview** model. * **Alibaba** opensourced the model **Qwen2.5 Coder 32B**, which offers comparable capabilities to leading proprietary language models in the coding domain. * **DeepSeek** unveiled its new AI model, **DeepSeek-R1-Lite-Preview**, which incorporates reasoning capabilities and delivers impressive performance on the **AIME** and **MATH** benchmarks, matching the level of OpenAI's **o1-preview**. * **Suno** upgraded its AIpowered music generator to **v4**, introducing new features and performance improvements. * **Mistral AI** launched the **Pixtral Large** model, a multimodal language model excelling in image recognition and advanced performance metrics. * **Google** introduced two experimental models, **gemini-exp-1114** and **gemini-exp-1121**, currently leading the arena chatbot with enhanced performance. source: [https://nhlocal.github.io/AiTimeline/](https://nhlocal.github.io/AiTimeline/)

Reply to Post

21 Comments

[-]

KeyObjective8745@reddit

Is it possible to get this news on rss? I would like to add it to nextcloud news

[-]

nh_local@reddit (OP)

I tried to work on it but got a bit stuck. You are welcome to contribute to the [repository on GitHub](https://github.com/NHLOCAL/AiTimeline). It is built with jekyll

[-]

jhymn@reddit

Is it just me or for those of us stuck using smaller localLLMs, is anyone extremely impressed with Mistral Nemo relative to its size? For me, as much as I try flirting with other models on my local machine, Nemo is the one I keep coming back to over the last couple of months.

[-]

s101c@reddit

Similar story happened yesterday. I switched to Mistral Small 22B when it came out and thought it was much more impressive than Nemo (and it was). Fast forward to yesterday, I got a new project. Parsing news for insights and compiling it into readable digest (my scrolling habits have become unbearable). It was the first time when Mistral Small's result (due to a low quant most likely) was not impressive at all. I decided to get Nemo because of the largest context window and wow. Analysis was on point, it provided more interesting digest than the other comparable models (and I have tested Gemma 9B, Llama 3.1 8B, 3.2 3B). And so, Nemo got a new life.

[-]

naaste@reddit

The recent developments in AI from companies like Mistral AI and DeepSeek indicate a highly competitive landscape. With benchmarks being established, especially in reasoning and multimodal processing, it raises questions about how these innovations will redefine our expectations of AI technology. For instance, how do you think these new models stack up against existing offerings from OpenAI? Delving into this could deepen our understanding of their potential advantages and challenges.

[-]

remghoost7@reddit

**That Suno v4 model is pretty rad.** The vocals are clean as heck. My only complaint is that the instrumentation is a bit compressed, but that could be cleaned up with mastering in a program like Ableton. [Here's one of the songs made with it.](https://x.com/imolivercom/status/1856488669747573051) Sorry, garbage twitter link. \--- **So glad to see people testing more "complex" musical styles** (Nu-Metal in this instance). It seems like everyone and their brother is making Country music with AI, which is so musically boring (in my opinion, obviously) and a horrid example of a model's quality. A good test of an AI music generation model is a genre that has intricate guitars/drums, melodic vocal lines, breakdowns / half-time to double-time transitions, etc. **It shows the depth of the model and what it's actually capable of.** \--- **That's one of the reasons I fell in love with the Udio model** (specifically the 1.0 model, not the 1.5 model). It can make super weird genres like ["melodic post-hardcore jrock"](https://www.udio.com/songs/nYByeRSRGhXbTE3STpiQJT) surprisingly well. Intricate guitars/drums, interesting melodies, odd rhythmic variations, double-time to half-time transitions, etc. **Having an AI music generation make Country music would be like asking an LLM to spell the word "cat".** Everyone has different tastes in music (and I completely respect that), but it's more or less a worthless example of what a model is actually capable of. \--- **It's just a shame that we'll never see a good one of these AI music models hit the locally hosted scene.** **Stable Diffusion came around so quick that no one knew what hit them.** And the art scene is too distracted with arguing about it that no regulation will ever reasonably come down on it. **LLMs are already so robust that future training data could be cut off entirely and we could just generate decent training data from the models we already have.** I believe most synthetic datasets are already generated via GPT4 nowadays. **Music models are screwed though.** The record labels are already keen to their music being used to train models and have the money/direction/incentive to make sure it never takes off. Closed source is the only way we'll probably ever get decent music models (essentially by companies hiding their datasets), and those companies have no reason to release those models. Still holding out for some disgruntled employee leaking a Udio/Suno model, but it's highly doubtful. **It's a damn shame.**

[-]

Dmitrygm1@reddit

I've tried it a bit and have mixed impressions. The vocals are a lot cleaner than previous versions, very impressive... however, the beats still sound distorted and have a distinct sound which immediately gives away it being AI. But occasionaly, it generates stuff which is geniunely mindblowing, like this one: [https://suno.com/song/4ec14116-ac9b-49e3-ae62-99eb2f1ac0c2](https://suno.com/song/4ec14116-ac9b-49e3-ae62-99eb2f1ac0c2) I actually would not be able to tell this is AI generated.

[-]

mwmercury@reddit

I love the way you put the news about closed source models in the last lines

[-]

nh_local@reddit (OP)

The truth is that it's not on purpose... there just haven't been many events of closed source companies

[-]

DeltaSqueezer@reddit

Why are some with blue dots and other red dots?

[-]

nh_local@reddit (OP)

Particularly significant and influential events that constitute a breakthrough in the industry are marked in red. Everything else is in blue

[-]

Kathane37@reddit

you could cerebras touching the 1000 token/sec with llama 3.1 405b Which would be amazing paired with reasoning model

[-]

bymechul@reddit

I don't know how cerebras did it, but it is very good for the future. but it's very expensive. 1m cost per token

[-]

mintybadgerme@reddit

I don't suppose there's any chance of adding live links to some of these announcements, because it'd be great to go straight to them from the main post?

[-]

nh_local@reddit (OP)

It will be quite a job to collect the sources. The truth is that I gathered a large part of the information from the amazing site [THE DECODER](https://the-decoder.com/)