How does a car benefit from having Deepseek installed ?
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
I can think of one, you have an accident, you get knocked out. The car sensing its upside down and something is not right or something calls emergency services and Deepseek takes over to get some help to you.
Big news! Qwen 3 is here!
A lot of models, all open sourced!
Among them an impressive 235B parameter MoE model (same model family like DeepSeek)
Trained on 36T tokens! That's impressive, the biggest token budget among opensource models (closed ones don't publish this info)
A lot of models, all open sourced!
Among them an impressive 235B parameter MoE model (same model family like DeepSeek)
Trained on 36T tokens! That's impressive, the biggest token budget among opensource models (closed ones don't publish this info)
Baba moving faster than deepseek. I think deepseek's edge is gone now that all the big chinese techs are moving aggressively on AI. Deepseek can only keep up if they get huge funding boost. Otherwise just having 200 really good employees will not cut it.
Qwen 3 scores are unreal. I mean the flagship model is better than o3-mini in reasoning, that's crazy.
But even more crazy is 4B model being as good as 72B model from Qwen2.5, which itself was as good as Llama3-70B model. So yeah, that's ridiculous.
Gemini 2.5 pro is still better though (not sure about o3 since it seems to perform better in bench marks than reality) and multi modal as well. Google is already starting to train their next model and no one has been able to beat them on context length. Competition is getting heated & winners are starting to emerge, still hoping for a Deep Seek upset because Google is looking invincible right now.
It's important to note that Qwen3 was never aiming for SOTA, since a 235B model can’t compete with models several times its size, such as o3 and 2.5-pro. They’re focusing on building extremely efficient models for local use. In that sense, DeepSeek and Qwen complement each other very well, with DeepSeek focusing on large SOTA models like V3 (currently the top non-reasoning model), while Qwen focuses on leading in small, efficient models that can be run locally.
It would be a waste of resources for Qwen to release a SOTA model that briefly surpasses o3/2.5-pro, only to be overtaken once R2 launches. This release proves that China is increasingly dominating efficiency at the 0.6B to 235B scale with Qwen3, and is now the undisputable leader in open-source AI across all model sizes. Qwen3 has made massive progress in efficiency and quite surpasses expectations.
It would be a waste of resources for Qwen to release a SOTA model that briefly surpasses o3/2.5-pro, only to be overtaken once R2 launches. This release proves that China is increasingly dominating efficiency at the 0.6B to 235B scale with Qwen3, and is now the undisputable leader in open-source AI across all model sizes. Qwen3 has made massive progress in efficiency and quite surpasses expectations.
Honestly it doesn't seem very productive just to spend loads of money making a model that will look good on the benchmarks for about two months and then it gets overtaken by someone else who releases a newer model. There needs to be more focus on utility.
They are not making this models "to look good" in benchmarks. For start the models are impressive for their size, from my tests they are pretty good. The models are made with efficiency in mind, Qwen3-30B-A3B is as good as a 70B model but can run in a CPU at a good speed. Qwen3-235B-A22B is as good as Claude and openAI <in my experience> but can fit 4bit quantized in 160GB RAM or 4 H100 unquantized and run at a decent speed. This models are made to be run as cheap as possible. Another thing that the Qwen team is focusing with this release is tool calling and agentic modes as these models seem to be pretty good at.
I dont think they have illusions that the model will be SOTA for long time that is why they release the models in a Apache license instead of the typical Qwen license because these models seem to be a snipped of what they want for future models, Cheap to run highly intelligent models with good agentic capabilities, although is possible that they will release a even bigger model for their paid chat.
The Qwen 3 coder is going to be a beast given how good this models perform