September: 2026 China AI Large Model Platform Ranking
Code Large Models Take the Lead in Entering the RSI Era
This wave of RSI (Recursive Self-Improvement) for large models was first ignited not in chat capabilities, but in code and systems. FaceMind Intelligence allowed AI engineers to write a complete pre-training framework from scratch, which took about 8 hours in total, with performance matching NVIDIA's open-source Megatron-LM; the Infra Agent disclosed by Zhipu in mid-September allows GLM-5.3 to locate the performance bottleneck of the GLM-5.3-Flash inference infrastructure, modify the underlying code, run layered tests, and iterate based on feedback.
It can be seen that the reason why code has become the first test field for RSI is that it is currently the only field where "the model can automatically verify right or wrong after finishing writing". Capabilities such as text, dialogue, and reasoning still require human scoring, but code allows the model to act as its own referee.
Although they are all talking about RSI, the puzzle pieces they cut into are actually different. Zhipu takes the path of "model optimizing the inference system", closing the first loop in the minimum RSI closed loop, that is, the model first modifies the infrastructure it relies on to run; the Qwen team's Qwen3.8-Max locates defects by itself based on internal test logs and user feedback, generates training data on its own, designs controlled experiments independently, and then hands it over to small models for multi-round verification to filter out invalid data. It ran 33 rounds of iterations unattended in one month, and the Artificial Analysis score rose from 40 to 45.
In addition, Unisound's U2-Flash released in September focuses on the post-training closed loop, specifically targeting those boundary tasks that "can be half done but are prone to errors and have long paths", directly converting failure trajectories into training materials, and using multi-teacher distillation for sub-domain verification; MiniMax's M2.7 in March this year follows another idea, allowing the model to build dozens of complex skills in the reinforcement learning Harness by itself, update its own memory, and even cooperate with universities to develop ModularRSI in September — the base weights are frozen throughout the process, and only the Agent is allowed to modify the peripheral operating mechanism. One modifies the inference stack, one modifies the training data, one modifies the post-training trajectory, and one modifies the Harness. Essentially, they are seizing positions in different links of RSI.
A more notable signal is that companies are no longer satisfied with "letting the model optimize tools" and have begun to move towards "letting the model train itself". At the semi-annual performance meeting at the end of August, Tang Jie of Zhipu directly defined GLM-6.0 as a Full Self-Training model; Alibaba clearly wrote RSI into the machine intelligence roadmap at the Yunqi Conference, and announced that the next generation will launch a brand new model with 5 to 10 trillion parameters.
In other words, the previous few companies are still "the model repairing the water pipes for itself", and the next step is to compete for the model to participate in building the next generation of itself. If this step is really completed, the main body of Scaling will no longer be only researchers, but the model itself rolling forward on the time axis.
It is still far from RSI in the real sense. Zhipu itself also admits that the system has not yet reached the stage of fully autonomously designing and training the next generation of models, and only the minimum closed loop is closed; most of the disclosures from various companies are "one link rotating independently", rather than "the improvement of the previous round can be completely inherited by the next round". In the five-level framework of Theseus Labs, most domestic teams currently stay at level 2 to 3.
The real watershed is not who shouts the word RSI first, but more about whether everyone can run this loop stably, quickly, and cheaply. After all, RSI is essentially Scaling on the time dimension. Now it is more like everyone has laid the first brick of RSI. The real recursive cycle will not be able to see who is really running until the next generation of base models comes out at the end of the year.
Personal Agent Stuns the AI Circle
The recent boom in Personal Agent is behind several large manufacturers pushing "personal AI assistants" a big step forward in a very short time.
On September 8, Meta launched Muse, which equips each user's agent with a dedicated virtual computer that can open browsers, fill out forms, process emails and arrange trips; after closing the application, it can continue to work, and when operations requiring approval such as purchasing and sending emails are needed, it will come back to the user for confirmation. Then on September 29, OpenAI launched dots at the developer conference, driven by GPT-6 Astra, which also has a cloud computer, can connect to external applications, and collaborate with users at entrances such as ChatGPT, Slack, and Teams.
The intuitive change brought by these two products is that after the user assigns a goal, the AI starts to take over the subsequent promotion work, and the user does not have to stay in the dialog box all the time.
Domestic manufacturers are also accelerating, but the entry points and progress of several parties are different. Qwen announced at the Yunqi Conference on September 22 that it would accelerate the creation of Personal Agent, focusing on continuously understanding needs by combining personal information provided by user authorization. For example, around a running goal, adjust training suggestions according to exercise records, extending a single consultation into long-term services.
Doubao is extending the capabilities accumulated by the mobile assistant to a wider range of personal assistant scenarios. According to reports, its exploration project codenamed Spell was internally tested as early as April, and is currently accelerating integration with Doubao products, which is still part of the product planning disclosed by the media. Ant Lingguang recently confirmed that the new version is in small-scale internal testing, and the use cases in public reports include managing DingTalk check-in, tracking market changes, and rearranging travel routes after flight delays. The common direction of several companies is already very clear: let AI take over things that require continuous follow-up and changes in the process.
In fact, Personal Agent can be understood as a personal agent that remembers your situation, can call tools, and can continuously advance tasks. "Personal" means that it knows your preferences, restrictions, and what you are doing, reducing the need to explain the background from scratch every time; "Agent" means that it can decompose tasks, operate applications, and adjust the next step according to the execution results. For example, to arrange a business trip, the user hopes that it will screen plans in combination with the calendar, budget and travel habits, complete the booking after obtaining authorization, and continue to follow up when the subsequent itinerary changes. This is also the most critical change from conversational AI to action AI: the end point of an interaction extends from receiving an answer to making progress and getting results. The memory, background operation and tool connection emphasized by Muse and dots are all completing this link.
From an industry perspective, this will further push the competition of model companies to shift to complete task delivery capabilities. Performing well in a single answer is two different tests from continuously executing dozens of steps and continuing to complete the task when encountering problems. The model needs to understand the goal, select tools, check results, and also judge when to stop and ask the user; the product system needs to handle memory, permissions, task status, and exception recovery.
Of course, in addition to the model ranking, there are also the completion rate of real tasks, how many times the user needs to take over, and how much time and computing power it takes to get one thing done. Model capabilities still determine the upper limit, and the system that stably organizes these capabilities will more and more directly determine the user experience.
At a deeper level, Personal Agent may also change the way large model products compete for users. In the past, users often switched between several AI applications according to the effect of a single answer; when an agent gradually gets familiar with your projects, schedules and work habits, changing products means re-establishing this understanding.
The competition among manufacturers will further extend to the long-term usage relationship of users and the ability to connect real services. However, this relationship must ultimately be established through reliable deliveries again and again. Whether Personal Agent can become a new growth point in the large model industry from the recent boom depends on whether users are willing to hand over the next thing to it after using it.
AMD Invests 8.2 Billion USD Heavily, Why Has the Robot "World Model" Become a Must-Win Battlefield?
At the end of September, AMD announced the acquisition of World Labs founded by Li Feifei for about 8.2 billion USD, making the world model once again the focus of the AI circle. This transaction is yet to be completed, but the identity of the buyer is already very meaningful: a chip company has begun to acquire model teams to understand what computing power the next generation of AI needs in advance. AMD explicitly mentioned robots, simulation and physical AI in the announcement. Turning to China, we will find that another group of companies are also accelerating their actions — manufacturers that have entered the public view with robot bodies such as Unitree, Agibot, and Songyan have recently intensively released their own "brains".
In the first half of September, Songyan successively launched the HERON series of models, Agibot released AGILE 2.0 and GE-Act 2.0, and Unitree open-sourced UnifoLM-WLA-1.0, and continued to open the base model and fine-tuning code at the end of the month. All these revolve around a question: when robots have more and more flexible bodies, how can they judge by themselves and complete tasks in unfamiliar environments? Among them, Agibot tried to put future state prediction and action learning into the same model, while Unitree emphasized using one model to coordinate multiple tasks and adapt to different grippers and dexterous hands. Ontology manufacturers have begun to deepen competition at the model level, striving to master the ability of robots to understand the environment and make decisions.
The world model has attracted attention precisely because it tries to make up for the "prediction" before action. For example, to let a robot pick up a glass of water, it needs to judge the position of the cup, and also consider where to grasp, how to move, and what may happen after tilting. The world model hopes to predict what will happen next by learning the relationship between the environment and actions, and then provide a basis for planning and execution.
Agibot GE-Act 2.0 tries to optimize the predicted future state together with the robot's actions, so that the "imagined result" and the "actual executed action" correspond better. Along this direction, the combination of the world model and VLA is worthy of continuous attention: whether the prediction ability can help the robot make more reliable actions will eventually be verified through the real machine tasks.
Who can continuously obtain valuable data, and who can turn this data into reusable capabilities. Every time a robot enters a new workstation, it may encounter new objects, placement methods and operation requirements. If a large amount of re-collection and re-training is required each time, the deployment cost will be difficult to reduce.
One of the expectations carried by the world model is to let robots accumulate a more general understanding of the physical environment and reduce repetitive work when adapting to new tasks. However, there is still an engineering distance between "being able to predict" and "being able to work stably". How the model recovers after object occlusion, grasping failure, and environmental changes will directly affect whether it can enter real business. For manufacturers, the ability to continuously verify and improve after the model is released will become more and more important.
The robotics industry is re-dividing the core links that each company needs to master around intelligent capabilities. Chip manufacturers hope to get closer to model evolution, and ontology manufacturers hope to precipitate their own data, hardware and control experience into the model.
Domestic Large Model Updates
Alibaba
On September 28, the Qwen App was deeply connected with Quark Cloud Drive. After user authorization, users can query, organize and read cloud drive materials in Qwen, and generate review tools, work summaries and photo collections accordingly. Quark Cloud Drive will become a data asset center shared by users and AI. Membership rights and interests are also integrated simultaneously: purchasing Qwen App Elite and above membership will give away Quark Cloud Drive membership of the same duration. After stacking education discounts, the dual-end monthly membership only costs 25 yuan.
At present, subscribing to Qwen App Elite, Flagship and Supreme membership will give away Quark Cloud Drive membership of the same duration and different levels. After on-campus college students and teachers complete authentication, they can obtain both Qwen Elite membership and Quark Cloud Drive VIP rights for as low as 25 yuan per month.
On September 23, Qwen officially released the full series of Qwen-Audio 3.1 speech large models, launching 5 products at one time: the three core basic models of Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Realtime Speech Interaction (Realtime) are upgraded, and the next-generation architecture ASR-Next audio understanding model and TTS-Next audio creation model are released at the same time. Among them, Qwen-Audio-3.1-ASR outperforms the mainstream open-source models of the same period in 6 of the 11 subsets of the KeSpeech and WSYue public dialect test sets, with an average Character Error Rate (CER) of only 4.55%, performing prominently in multi-dialect and low signal-to-noise ratio scenarios.
On September 22, Alibaba announced a series of progress on large models at the Yunqi Conference. Recursive Self-Improvement (RSI) of large models has initially entered the links of model training, inference and chip-model collaboration. Qwen4 based on the new generation architecture is already under training, and future versions such as Qwen4.5 and Qwen5 will be expanded to 5-10T parameters. At the same time, video generation models, speech models, image models, world models, music models and full-modal models will all launch new versions on the same day or in the near future.
On the same day, the new generation of full-series Qwen AI hardware made its first concentrated appearance at the Yunqi Conference, including Qwen AI Glasses N1, N1 Pro and Qwen AI Ear-clip Headphones. The three new products opened online reservations on the same day and will be sold on October 13; based on the cognition that "data is the most important prerequisite for Agent to enter enterprise business processes", Qwen Office launched a new product for enterprises: Enterprise Context. Based on this product and the Agent hosting capability of Qwen Office, Qwen Office can help enterprises build digital employees who understand business, let them collaborate with human employees on various IMs such as DingTalk, reshape business processes, and greatly improve organizational efficiency; Wu Yongming, CEO of Alibaba Group, said that machines are becoming the main force of thinking, and the total amount of machine thinking in the future will be 1000 times that of humans. Today, this proportion is less than 3%, and there is at least tens of thousands of times growth space. AI models, AI chips and AI cloud are the three cornerstones of the era of machine intelligence. Alibaba will firmly invest in the construction of these three infrastructures, which is Alibaba's long-term strategic choice.
In terms of AI models, the Qwen team is exploring Recursive Self-Improvement (RSI) and plans to train a new model with a parameter scale of 5-10T. In terms of AI chips, Pingtouge released the new generation AI chip Zhenwu V900, whose computing power is 3 times that of Zhenwu M8900, and a single cluster built on it can be expanded to 500,000 cards. In terms of AI cloud, Alibaba Cloud aims to have a global data center scale of more than 20GW by 2032.
On September 17, Qwen added access to the People's Medical Publishing House professional medical knowledge base and NutriData nutrition database, further expanding medical and health professional capabilities to medical knowledge and nutrition data. When users ask health questions such as physical examination indicators, disease knowledge, and daily diet, Qwen can refer to authoritative medical knowledge and specific food composition data to provide more evidence-based and targeted health interpretations.
The People's Medical Publishing House professional medical knowledge base brings together the authoritative medical resources accumulated by PMPH for more than 70 years, covering core knowledge such as diseases, examinations and tests, and drugs; NutriData is co-built by the Chinese Nutrition Society, based on authoritative data such as "Chinese Food Composition Table". Qwen is continuously accessing authoritative professional knowledge bases, taking the "professional knowledge base" as a long-term basic capability in the medical and health field.
On September 11, Alibaba Cloud Token Plan personal edition was upgraded. The original price and Credit quota remain unchanged. The Standard and Pro packages add new Agent Harness tool rights and usage, covering 12 common Agent development capabilities such as search, web page parsing, image generation, speech processing, and code execution. All the above tools are open through the M