It's the first time in the history of the world.
The more models there are, the more the entrances are not necessarily more useful.
On July 29, Agnes AI officially launched a new domestic official domain name:
At the same time, the API access gateway has completed the switch, with Agnes 2.5 Pro Alpha taking the top spot in the list of Artificial Analysis Comprehensive capabilities, and 2.5 Flash continuing to be open for free indefinitely.
A battle has been waged over model capabilities, developer inputs and multi-modal landings.
Internal access to entry upgradedSynchronization of API servicesAs the AI application ecosystem continues to expand, stable access to and development experience for model services has become a major focus for developers.
The launch of Agnes' domestic official website is an important step in further improving its domestic user service system. In order to accommodate domestic access input adjustments, Agnes has also simultaneously provided new API service addresses.
Users who register with the original international station do not need to go to the domestic station to re-register, nor do they need to replace the original API Key.
Mainland Chinese users will need to continue using the original International Station account and the original API Key by modifying the original Endpoint interface to the following address:
In addition to the API address, the original API Key, model name, request parameters, and call method need not be modified. Once the address is replaced, applications or services can be restarted depending on the operating environment.
API platforms:
Artificial Analysis in the testAgnes 2.5 Pro Alpha tops the list of comprehensive capabilitiesThe results of the first round of evaluations of Agnes-2.5-Pro Alpha on Artificial Analysis have also been officially released.
According to a list published by Artificial Analysis, the model performs more prominently in indicators related to integrated intelligence, coding and terminal tasks.
Agnes 2.5 Pro Alpha's Intelligence Index scores 39 and ranks 9th out of 153 model reviews.
The index combines multiple dimensions such as real-world job tasks, Agent tool use, terminal coding, scientific programming, complex reasoning, and context, bringing the model closer to a comprehensive performance in a real work environment rather than a single knowledge-based Q&A capability.
Coding Index scores 58.8 and Terminal-Bench v2.1 scores 67.0%.
Link to the list:
Among them, Terminal-Bench v2.1 requires the model to enter the real terminal environment, execute commands, read and modify files, and complete software engineering and system tasks, in close proximity to the actual use of the current AI Coding Agent.
Flash is freePro faces complex development scenariosIn addition to the runway, the positioning of the Agnes 2.5 series is also clearer.
Agnes 2.5 Flash is designed for everyday coding, Agent and high frequency development tasks, suitable for handling documents, generating code, fixing bugs and modifying projects, and more importantly, Agnes 2.5 Flash will continue to be open for free indefinitely.
This means that a large amount of daily development needs can continue to be completed through Agnes's model without having to be hampered by cost barriers.
As an early version of the Pro series, it has the core capabilities of its flagship model, targeting complex Agent, professional coding and high-difficulty engineering tasks.
Instead of Flash, Pro Alpha focuses more on understanding and reasoning complex problems, execution of long-link tasks, collaboration on large projects, and complex code analysis and engineering development.
At the same time, the Agnes 2.5 Pro Alpha supports 1M ultra-long context windows, which can handle larger code warehouses and business data.
Two missionsFrom Interactive Games to Single File Web ToolsTo observe the performance of Agnes-2.5-Pro Alpha in actual development tasks, we set up two tests, the first of which was content generation and interactive organization, and the second test was more specific algorithm and tool development.
The first task was to get Agnes 2.5 Pro Alpha to create a scripted interactive game web page.
Tip words do not provide a complete script, nor do they specify the details of the characters, plot and technical structure, requiring only models to refer to the common play of text interactive games, designing their own stories and interactions.
In order to quickly verify the gameplay, character images and video material are temporarily reserved, prioritizing the completion of a functional game prototype, followed by the gradual addition of formal artistic resources.
Under this relatively broad demand, the Agnes-2.5-Pro Alpha adds story themes, character settings, scene arrangements, visual novel interfaces, attribute systems, plot branches and end pages, and integrates these modules into a web prototype that can run directly in the browser.
The second task was to generate an image-imaging tool called Color Echo.
The initial requirement is simple, uploading a reference image and an image to be processed, using the LAB color migration method, extracting color features from the reference image, and moving to another image. The entire processing process is done locally in the browser, without the need to build a backend.
After running the basic color simulation function, I gradually added a few rounds of real-time parameter adjustment, front and rear slider contrast, skin color and high-light protection, preserve formula and export high-resolution images.
Finally, in order to make the page more complete, I accessed a commonly used front-end web page to generate Skill, with a uniform optimization of the overall layout and visual style.
Eventually, the model integrates features such as color simulation algorithms, parameter controls, status saving, and image export into a single file web page.
This case is not a one-time generating completion, but a step-by-step refinement of the details through repeated modifications after the basic functionality is available, which is closer to the actual process of everyday product development and iteration.
Two examples show how the model performs in specific tasks, returning to the platform as a whole, and new changes to the call data recently released by Agnes.
According to data disclosed by Agnes, its full-mode model has reached 8.16 trillion weekly tokens, with 5.10 trillion text tokens accounting for 62.5% of the total, and 3.06 trillion multi-mode tokens accounting for 37.5%.[3]
In terms of call structure, text is still the main type of use, but multi-modal tasks such as images, videos have contributed more than one-third of Token processing, and the platform's use scene is expanding further into multi-modal.
In addition to the model, the Agnes team is also continuing to participate in the construction of the SGLang open source community.
According to GitHub's public records, the team's four Pull Requests have now been merged, and related work involves multiple card parallelism, cache optimization, cache token statistics, and KV Cache stability, primarily to improve the operational efficiency and system reliability of large model reasoning services.
As Agnes 2.5 Pro Alpha enters the developer testing phase, the official release of Agnes 2.5 Pro should not be too far away.