How can big open source models outperform commercial products?

To be honest, I was really surprised to see the performance of big open source models like Ali Tongyi Qianqi Qwen3. They not only catch up with, but in some ways even surpass commercial products. The reason for this is the unique collaboration model of the open source community - excellent developers from all over the world can work together to polish a model, and this kind of "hundred schools of thought" ecology is difficult for any commercial company to reach. And the loose agreement like Apache 2.0 makes this advantage more virtuous cycle.

I remember when Llama2 was released last year, the industry was still discussing whether the open source model could catch up with GPT-4. now it seems that this is simply a false proposition. Qwen3, for example, not only in the parameter scale (235B) far more than many commercial products, more critical is the breakthrough in the actual deployment efficiency - 4 H20 cards can run a full-blooded version, which is a boon for small and medium-sized enterprises. This reminds me of a few days ago to chat with a startup CTO, they are moving their business from GPT to their own deployment of Qwen, the cost directly down 70%, and the effect is basically equivalent. Is this not the best proof of the advantages of open source?

The speed of iteration of open source models is staggering

Commercial products are constrained by R&D processes and marketing strategies, and update cycles tend to be longer. But open source models are completely different, like Qwen's evolution from 2.5 to 3 in just a few months. What's even more amazing is the magnitude of performance improvement - Qwen3-4B can reach the level of its predecessor, version 72B. Behind this speed of iteration is the open source community's characteristic of "rapid trial and error, rapid optimization". When commercial companies are still going through the product review process, the open source community may have already tried more than a dozen improvement options, from which the optimal solution is selected.

There is another point that is often overlooked: data diversity. 36 trillion tokens of training data covering 119 languages, this scale and data diversity, commercial products tend to trade-off for cost considerations. However, open source projects can pool global resources, and different types of organizations contribute data from different fields. For example, there are research institutions specializing in mathematical data, some teams specialize in code generation, and ultimately fusion of an "all-rounder".

Hybrid Reasoning: The Wisdom of the Open Source Community

Qwen3's mixed reasoning mode is a genius design. Turning on deep thinking mode for questions that require deep thought, and quick responses for simple questions - this flexibility perfectly balances effectiveness and efficiency. Think about it, even commercial products can rarely achieve such intelligent mode switching. This comes from the pragmatic spirit of the open source community to "solve real problems".

What's more, the open source model provides a rich customization interface. Commercial products are often "one size fits all" experience, but Qwen3 not only allows to adjust the depth of thinking, but also control the thinking budget (number of tokens). This gives developers plenty of room to play, so that a basic model can be derived from countless "variants" to adapt to different scenarios. I am afraid that this kind of "building block" flexibility can never be provided by closed-source commercial products.

From the success of Qwen3, we can see that the future of the open source model is not to "catch up with commercial products", but to open up their own path driven by the collaborative ecosystem. When the commercial model is still entangled in the parameters of the scale and profit margins, the open source community has been in the actual deployment of the problem, the efficiency of the problem, the problem of personalization. This may be why more and more enterprises are turning to open source solutions - because in many practical scenarios, open source is not only enough, but also better.