The three giants in Silicon Valley have all pulled out their big guns. Now we're just waiting for Liang Wenfeng.

The battle among the titans in the Silicon Valley AI circle this summer officially kicked off on the evening of the 5th. OpenAI finally embraced open - source again after GPT - 2 and released a so - called "strongest" open - source inference large - model gpt - oss.

Google wasn't idle either. It directly launched a revolutionary weapon - Genie 3. It allows users to generate an interactive three - dimensional virtual world lasting several minutes with just one sentence, and the effect is astonishing.

OpenAI's old rival, Anthropic, which focuses solely on the AI programming track regardless of AGI, was also eager to make its mark. It updated its top - tier large - model: Claude Opus 4.1, further raising the upper limit of AI programming capabilities.

Although the releases of these three new products are of great significance in different aspects, this is just the prelude to the "battle among the gods" in the Silicon Valley AI circle in the coming days. The real show is yet to come.

Just like all the previous products released by the big three, Chinese people still play a crucial role in the development teams. So, even the onlookers in Silicon Valley are looking forward to the arrival of DeepSeek and Qwen from the East. They hope that the domestic AI forces won't miss this AI feast at the end of summer.

01 OpenAI Finally Reached Its DeepSeek Moment

OpenAI launched its "open - weight" large - language models, gpt - oss - 120b and gpt - oss - 20b, for the first time in six years. Both models adopt the Transformer architecture and incorporate the MoE design. The gpt - oss - 120b has a total of 117 billion parameters, with 5.1 billion active parameters. The gpt - oss - 20b has 21 billion total parameters, and 3.6 billion parameters are activated per token.

The models use the grouped multi - query attention mechanism with a group size of 8 and the Rotary Position Encoding (RoPE), natively supporting a 128k context.

In terms of performance, OpenAI officially claims that these models are the state - of - the - art (SOTA) among open - source inference models of the same scale:

The gpt - oss - 120b model achieves performance close to that of OpenAI o4 - mini in core inference benchmark tests and can run efficiently on a single 80 GB GPU. The gpt - oss - 20b can achieve performance similar to that of OpenAI o3 - mini in common benchmark tests and can run on edge devices with only 16 GB of memory.

In terms of model scale, the two models released by OpenAI this time are targeted at the demand and market for local deployment, filling the previous gaps in product offerings. Moreover, the license still has some tricky restrictions. Commercial use by entities with an annual revenue of over $100 million or a daily active user base of over one million is not allowed.

After the release, netizens couldn't wait to deploy the models locally, and the performance feedback has been good.

When running the 20B version on an RTX5090, the output speed can reach 160 - 180 tokens per second.

In terms of model capabilities, users' actual feedback has also been positive. This netizen passed three common programming ability tests at once on an M4 MacBook.

According to the crowd - testing results released by the large - model ability testing institution Intelligence Evaluation, the performance of OpenAI's two models is close to that of the two Chinese open - source models, DeepSeek R1 and Qwen 3, despite having a much smaller scale.

Objectively speaking, the most significant aspect of this model release is that it brings OpenAI back to the open - source arena. It gives users another option for a model that is easy to deploy locally on a single machine and has decent capabilities. However, it can only be regarded as a make - up release. Everyone's expectations for OpenAI still lie in the upcoming GPT - 5. Let's wait and see how much of a shock OpenAI can bring to the industry.

02 Google Genie 3: Astonishing, but a Future Promise

Although OpenAI didn't release GPT - 5, Google wouldn't allow OpenAI to monopolize the limelight. It released a very "futuristic" model - Genie - 3.

As shown in the video at the beginning, this is a model for generating virtual worlds from text. It allows users to generate a world model for real - time interaction in a dynamic world with a resolution of 720p and a frame rate of 24 FPS. It has excellent consistency and realism. Users can control and explore various details in this world using text or buttons.

 Caption: Walk to an aluminum rack, then walk to the large red industrial mixer.

Put simply, if text - to - video models like Sora or Veo 3 are equivalent to movies, Genie - 3 generates a game or a virtual script - killing experience. Users not only watch the world on the screen but can also actively interact with the generated world through their actions.

Different from video models, such a "world model" needs to provide real - time feedback and interaction based on the user's movements and must conform to the basic logical laws of the real world. For example, if you push a floating balloon backward with your hand, the balloon should fly backward instead of upward. If you gently push a car, it shouldn't suddenly speed off.

So, if users can tolerate some minor flaws in video models, a world model must handle physical feedback very precisely. Although it may not reach the level of "understanding the physical world" required by LeCun, compared to the fleeting images generated by video models, a virtual world that fails to handle these physical interaction logics is a meaningless product.

In the demo provided by Google, as users input different instructions, the world provides real - time feedback and generates different content. Achieving this requires addressing very challenging technical issues.

If Google can successfully implement this in the future, regardless of whether the technology behind the "world model" will bring about revolutionary changes in the fields of robotics and autonomous driving, its impact on the VR, gaming, and text - to - video industries will be unimaginable.

Unfortunately, although Genie 3 looks amazing, it is still in the official demonstration stage. All the released content is provided by Google, and no trial for ordinary users is available. It's still the classic Google - style future promise.

However, even though it's a future promise, currently, only Google, which has made a saturated investment in AI, seems capable of achieving such a revolutionary product and technological breakthrough.

Technically, even a startup like OpenAI can only compete with Google in the Transformer - based models. Although it was OpenAI that launched Sora over a year ago, which amazed the world and raised people's expectations for video - generation models, Google and its Veo are the ones with the real strength to make continuous investments and maintain a leading position over time. Genie 3 is also the result of Google's years of silent efforts and multiple major version iterations.

Considering that the Transformer architecture also originated from Google, we hope it will nurture every spark and let us feel the heat of the spreading fire soon.

03 The Persistence of the One - Track Champion: A 2% Improvement

One picture sums up Anthropic's release: the upper limit of AI programming capabilities has been raised by 2%.

It should be emphasized that this 2% doesn't just represent the improvement of Claude itself but the current upper limit of AI programming capabilities.

Judging from users' real feedback and market share, Claude Opus is almost the most well - regarded and widely - used model in AI coding at present.

AI coding is almost the most commercially successful and has the highest future potential among the sub - tracks of large - model applications. So, in the face of OpenAI's dominance, Anthropic has chosen to invest all its resources and energy in improving the programming capabilities of its model. Through this competitive strategy, it hopes to stay in the game of large - model commercialization and retain the possibility of continuous confrontation with Google and OpenAI.

From this perspective, the release of Opus 4.1 seems to be a firm assurance to all its customers, saying, "Don't worry