Summary of Key Points
Google has been updating its Gemini Flash large model every three weeks (the latest version 3.8 was released just three weeks after the previous one), and it has also introduced a Cyber version specifically designed for cybersecurity. Meta has followed suit by updating its own model and teased Google about "who Gemini really is." Google's flagship Gemini Pro series has faced difficulties in development (version 3.5 Pro was canceled, and version 4 is still in training), which has forced the smaller Flash model to take on a more prominent role. The reason behind this change is that the approach to large model development has shifted: instead of simply focusing on scale, there is now a greater emphasis on post-training techniques such as reinforcement learning and agents. The smaller size of the Flash model allows for faster iteration of these new capabilities.
1. Why Does Google Update Flash Every Three Weeks? The Smaller Model Has Become a “Fast Testing Ground”
In the past, large model updates were slow due to their massive size, which required significant computational resources, time, and cross-team coordination, resulting in high costs for experimentation. However, the Flash model is different—it is smaller, making modifications and retraining much less expensive. This enables multiple teams within Google to simultaneously test various techniques, such as reinforcement learning and agent training.
For example, the Flash model is like a “test version of the software” on a mobile device, allowing for quick addition of new features and bug fixes, whereas the Pro series is the “official, flagship version” that can only be updated when the features are fully mature. Now, Google tests new technologies like reinforcement learning and tool invocation on the Flash model first, and if they prove effective, they are immediately implemented, which is why Flash is being updated so frequently.
2. What Makes the 3.8 Flash Model Strong? It’s Like an “Efficient Worker”
Previous AI models would provide only preliminary results when performing complex tasks, without checking if the tasks could be executed correctly or if there were any errors. The 3.8 Flash model includes a new feature called “Agentic loops,” which means it can actively iterate through the process: it first determines how to complete the task, invokes the necessary tools, checks the results, makes adjustments, and repeats the process until the task is completed.
For instance, in an official demo, the model was given a simple instruction, and it was able to create a playable 3D wizard game with its own tools, including puzzles and texture resources, as well as a DOS version of Google Maps with street views. The model’s performance in long-term software engineering tests surpassed that of many other large models, and it showed significant improvements in financial and legal scenarios. Google aims to move away from the perception of Flash as a cheap and fast but less capable model.
3. Speed Is Great, but Are the Details Enough? Both Advantages and Disadvantages Are Clear
The advantages of the Flash model are its speed and cost-effectiveness:
- Compared to the Claude Opus 5 model, the Flash model costs only $0.12 to generate a similar Three.js scene (Opus costs $1.86, 15 times more), and it is also much faster (37 seconds vs. 24 minutes).
However, it has some drawbacks:
- When generating the Sticky Ball game, Flash produces a quick result, but the experience is not as good as that of the Kimi K3 model (there are differences in the movement mechanics and design).
- When creating a New York city scene, Flash only included 6 trees (Opus included 1840 trees), and it omitted the pedestrian crossings and underwater effects that were requested by users, while adding unnecessary camera presets and a real-time clock.
In summary, the Flash model is suitable for quickly completing tasks, but its precision and final quality are still inferior to those of the flagship models.
4. Why Has the Flagship Pro Series Faced Challenges, and Why Has Flash Become the “Emergency Solution”?
Google’s flagship Pro series has encountered difficulties:
- The planned version 3.5 Pro was canceled because it did not offer significant improvements compared to the Flash model.
- Version 4 of Gemini is still in training and not due for release for a while.
The reason for these challenges is that large models require substantial computational resources for each update, and only when there is a qualitative leap in their capabilities is it worthwhile to invest. The Flash model, with its smaller size, can iterate quickly, filling the gap left by the Pro series. Additionally, due to internal adjustments at Google (with management urging for faster development), the Flash model has had to take on a more central role, handling both cost and speed concerns, as well as the complex tasks that the Pro series was responsible for.
5. What Is the Cyber Version? A “Security Expert” Designed to Identify Vulnerabilities
The Gemini 3.8 Flash Cyber version is specifically designed for cybersecurity tasks:
- It can identify vulnerabilities, perform penetration testing, and automatically generate security patches.
- Test results show that it outperforms the previous version and many other large models in vulnerability detection benchmarks, with a high patching success rate compared to the best models, yet at a much lower cost.
- In practical use, the Chrome team used it to generate correct patches 2.6 times more efficiently than commercial models. A security company, Wiz, found serious vulnerabilities that would take months to detect using traditional methods in just under two hours.
This version is intended for trained cybersecurity professionals, providing them with a highly efficient and cost-effective assistant.
Conclusion: Is Google’s “Circuitous Rescue Strategy” Successful?
Google may not have intended for the Flash model to become the main focus, but the challenges with the Pro series and the shift in large model development strategies have made it a popular choice. However, this is only a temporary solution. Ultimately, the success of the Gemini 4 model will determine whether the Flash model can truly fill the gap left by the flagship series. Meta’s teasing also highlights that Google’s lag in flagship model development has attracted attention from its competitors.