The true level of domestic RSI: No one dares actually modify model weights, but everyone is changing the "machine for building AI." Author | Gao Yunyi
Editor | Cen Feng
RSI, or AI building AI, has become an internal scorecard that model companies compete to show off in 2026, and one of the hottest directions in China's primary market. As long as a BP includes the word "Self-Training," its valuation is at least 20 percent higher.
But true RSI is not just AI improving AI; it must also distill experience and allow the learning method itself to evolve.
In September this year, 35 researchers from institutions including Shanghai Jiao Tong University, Tsinghua University, ByteDance, and SenseTime published a paper titled "The Last AI Built by Humans," setting a benchmark for the entire industry. It divides RSI into five levels, L1 through L5, and its conclusions are sobering:
The systems of most leading global companies are at the transitional stage between L1 and L2. L3 and L4 have only begun to emerge in scattered, embryonic form in specific vertical domains, such as software engineering and scientific discovery, while a true L5 closed loop exists only in a handful of constrained laboratory prototypes.
If we use this yardstick to measure the RSI experimental results unveiled by domestic model vendors in September and October, includingtraining pipelines, inference systems, GPU kernels, and Agent scaffolding;butcurrently, not a singlevendorhas dared to truly use AI to rewrite the underlying model weights.In other words, everyone is still far from the legendary “AI inventing the next generation of AI on its own (L5)”; they have not even reached the threshold.
So, amid this RSI boom, which ecosystem niches have China’s major tech companies and prominent startups each settled into? What, specifically, are they using AI to improve, and how far have they progressed? Below, we survey the latest RSI status of domestic model vendors.
01
Zhipu AI RSI: Running on 100,000 domestic accelerator cards,
an inference system capable of self-optimization
Zhipu AI is the first company in China to truly run RSI in a production environment. Its first RSI step did not touch the model weights; it focused first on the dirtiest, most labor-intensive underlyinginference system. Once that was working, Zhipu AI turned its attention to the Agent’s Harness. The former cuts costs, while the latter boosts efficiency—both are substantive, hands-on engineering work.
Start with the inference infrastructure. For a new model to actually become usable by users, a great deal of dirty, hard work lies in between: how to squeeze every ounce of computing power from the chips, how to split massive workloads across tens of thousands of cards to run simultaneously, how to efficiently manage the enormous volumes of temporary data generated during conversations, and so on. In the past, these tasks were tuned line by line by Infra engineers.
This time, Zhipu AI directly handed all of this over to the “Infra Agent” powered by GLM-5.3. On September 17, Tang Jie published a long post on X titled“The Path to Recursive Self-Improvement”, stating that the GLM-5.3-powered Infra Agent had built a production-grade inference service for GLM-5.3-Flash from scratch in just two weeks on a cluster composed of 100,000 domestic chips.
While handling all production traffic, this system pushed end-to-end throughput to 3.2 times the initial baseline, bringing hardware utilization and cost per Token in line with the levels of mainstream NVIDIA GPUs. With this self-built inference stack, Zhipu anonymously launched GLM-5.3-Flash, which processed more than 62 trillion Tokens in six days.
These proven optimization solutions will be abstracted into an “optimization skeleton” and stored in an asset library for direct reuse the next time a similar bottleneck arises. In Tang Jie’s words:“Zhipu has achieved a minimal closed loop: models optimize systems, and systems serve models.”
On the day the news broke, Zhipu’s Hong Kong-listed shares rose 8.29%, and the market readily paid up for the results of “Zhipu RSI.”
Shortly afterward, in late October, the Tsinghua and Zhipu teams released an updated paper《HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs》, with Tang Jie personally listed as an author and its sights set on the Harness for Agents.
As the orchestration framework surrounding large models, the Harness oversees prompt design, tool invocation, compute-budget allocation, and the flow of task routing. Previously, this intricate set of control parameters could only be tuned manually by engineers. If AI were left to modify them itself, it would be especially prone to “peeking at the ground-truth answers” and cheating, producing a high-scoring but underperforming system.
HarnessPro’s breakthrough lies in the fact that it does not touch the ground-truth answers at all. The system deploys multiple specialized Agents to collaborate: a diagnostic Agent identifies the bottleneck, a test-generation Agent synthesizes test cases itself for stress testing, and a planning-and-implementation Agent modifies the framework’s code according to different objectives—saving money or improving performance—while running “accuracy” and “execution cost” together as two metrics.
As a result, without any exposure to the answers, the pass rate on high-difficulty coding tasks rose rather than fell, and per-task spending was slashed by 53%.
This series of moves precisely validates the industry maxim from Zhipu founder and chief scientist Tang Jie: “What drives the loop is not intelligence, but the closed loop itself.” Even if the model itself is not a genius, as long as there is something that can be improved, a verifiable judge, and a sustainable feedback loop and these three are interlocked into a single closed loop, performance can be pushed to its limit.
Clearly, Zhipu’s ambitions do not end here. At the semiannual results briefing on August 31, Tang Jie had already explicitly defined the next-generation GLM-6.0 as a “fully self-trained” model and planned to pour roughly 60% of the net proceeds from the US$5 billion it had just raised into this self-evolution system.
In a speech in Singapore in early October, he again laid out three routes for GLM-5.4 / 5.5: scale the pretraining data to several trillion tokens, meaning the parameter count could jump from the current 74 billion-plus to at least 1 trillion; strengthen RL post-training by taking SFT, RLHF, and RLVR further; and continue to strengthen long-horizon reasoning and agent capabilities.
He emphasized that AI now basically no longer needs to rack up conversations or write code with Coding. The next generation will take two paths: RSI autonomous iterative upgrades and Agents with longer cycles and greater autonomy.
But he did not shy away from the biggest challenge either. “It’s not about continuing to scale the model,” he said, “but about enabling the model to reliably determine when to stop, when to correct errors, and how to evaluate the results of its own training.”
02
DeepSeek RSI: DSec Sandbox Factory, 3 Million “Testbeds” a Day
DeepSeek’s RSI is unlike that of its peers. The first thing it redesigned was the Agent’s training ground—the sandbox.
RSI is essentially high-density trial and error. Once one generation of Agents has been trained, engineers still have to build the training environment for the next round from scratch. Unless this bottleneck is resolved, no matter how intelligent the model is, it cannot get up and running.
To break this deadlock, DeepSeek and Tsinghua University jointly launched the automated sandbox cluster DSec on September 19.
It is a true “AI testbed factory”: it has a massive infrastructure of 30,000 CPU cores and 250 TB of memory and can produce around 3 million sandboxes per day. From lightweight function calls to full virtual machines, its four tiers of isolated environments respond within seconds and are available on demand.DSec.
Here, AI took control of the pace of its own evolution for the first time. No longer dependent on manually built training grounds, it formed a systemic cycle of “Agents automatically building environments—the environment gives rise to stronger Agents—stronger Agents feed back into better environments,” and all these environments can be reused as assets.Agent automatically builds the environment—the environment gives rise to stronger Agents—stronger Agents feed back into better environments.
Meanwhile, DeepSeek is also improving Harness.
The DeepSeek Harness, open-sourced in August, is built around the idea of “everything is a plugin.” It modularizes the agent’s tools, logs, and even core runtime framework. The benefit is that Agents can “upgrade while running,” and even if a change goes wrong, it can be rolled back with one click.Previously, upgrading an AI framework required shutting down and restarting; otherwise, the system would become increasingly chaotic with every change. Now that it is modularized, AI can uninstall and replace components at any time without downtime, in a highly secure manner.
To provide solid technical support for this “plug-and-play at any time” capability, DeepSeek and Peking University elaborated in detail on the underlying Cordis mechanism in the paper “A Programming Paradigm for Spatiotemporal Composability”. This mechanism was purpose-built to solve two long-standing industry pain points: first, how to roll back to the original state if something goes wrong; second, how to automatically adjust dependencies when a plugin is removed without affecting other modules.
This design has been running stably on the open-source chatbot framework Koishi for four years, successfully supporting more than 4,000 community plugins.
On October 21, DeepSeek also launched DeepEvolve, a self-evolving agent capable of “predicting the future” DeepEvolveTAG380@@, a critical step in RSI (recursive self-improvement): enabling AI to learn how to evolve itself in the real world, where information is polluted.
TAG388@@Its core is to decouple “information retrieval” from “result prediction.” By precisely distilling reusable predictive skills from limited real-time trajectories, even when external information is chaotic, AI can filter out distractions, distill genuine insights through retrospective analysis, and achieve closed-loop algorithmic iteration.
Every move Liang Wenfeng makes may appear scattered, but in reality, they all point toward the same endgame.
From the MoE architecture and MLA attention mechanism to the DSA sparse algorithm and now the DSec sandbox and self-reconfigurable Harness, every DeepSeek innovation is doing the same thing: lowering the computational barriers to AGI and RSI.
03
ModelBest RSI: An Automated Factory for “Building AI”
ModelBest’s RSI is about having AI write the low-level foundational software needed to “build AI.” In a typical AI-chip adaptation process, especially when targeting domestically produced chips, human engineers must perform extensive low-level operator tuning, rewrite communication operators, and coordinate software and hardware allocation. This is grueling work that relies heavily on experience and takes months, directly determining whether the chip’s computing power can be fully utilized. ModelBest has turned the iterative development of this low-level adaptation code over to AI. The performance of AI-generated code can reach up to 1.6 to 5 times that of code handwritten by top-tier engineers, and the months-long software-hardware adaptation cycle has been compressed to a matter of days.
In May this year, ModelBest internally launched the “Forge Engineering” project. Li Yuxuan, ModelBest’s chief scientist, summarized the project’s core idea as: “Provide more environments and fewer instructions.” The entire project advanced step by step along the underlying software stack for building AI.
The first stepwas to develop the training software ForgeTrain.On September 24, ModelBest announced the results: the AI had independently written the large-model pretraining framework ForgeTrain from scratch, matching the benchmark performance of NVIDIA’s mature Megatron-LM framework in just eight hours and surpassing its performance within 2.5 days. If this work had been handed to humans, it would have required at least three to five top engineers and more than six months.
The second stepwas to develop the low-level operators ForgeKernel.On October 27, ModelBest announced its latest progress: the AI had descended even further, to the lowest level of the chip, to rewrite the “last mile” connecting models and hardware. This self-iterating ForgeKernel operator system took first place on the official operator leaderboards of both Huawei and NVIDIA. In tests on the Ascend 950 series, its benchmark speedup reached 1.6 to 5 times that of the runner-up, effectively improving compute utilization without increasing hardware investment.
The third stepwas to rewrite scientific computing. The AI began crossing domains to take over the writing of some low-level core code for real-world industrial simulation and scientific computing.
From training frameworks and low-level operators to scientific computing, the AI gradually took on this kind of low-level code development work. Meanwhile, human engineers shifted to establishing standards and evaluation metrics (Harness).
Worth noting, ModelBest was also one of the institutions deeply involved in the September report “The Last AI Built by Humanity.” If ModelBest’s project is viewed within this roadmap, it has reached the stage where AI is deeply involved in the engineering implementation of “AI manufacturing AI,” but has not yet reached the top-level decision-making stage of “deciding for itself what to train and why.”
04
Moonshot AI RSI:
Autonomous Iteration on Long-Horizon Tasks
Moonshot AI’s RSI follows a “mastery-first” approach. It invested substantial computing power in large-scale reinforcement learning, working aggressively to enhance the AI’s long-horizon autonomous capabilities.
During the post-training phase, Moonshot AI used KDA hybrid linear attention and the MoonEP training infrastructure to enable the model to iterate repeatedly in real-world task environments such as human simulation, finance, and 3D modeling.
On July 16, Moonshot AI released and open-sourced Kimi K3, with a total parameter count of 2.8 trillion. To support such high-intensity agent training, the team built a sandbox system called AgentEnv. During training and evaluation, it created more than 51 million sandboxes and used over 1.5 million images. However, under extreme stress tests, traditional containers frequently experienced kernel crashes and deadlocks, often interrupting training. The team ultimately switched to Firecracker microVMs as the technical backbone.
If you break this system down, Moonshot AI has actually implemented automation at three levels.
At the task level, humans provide only a sandbox environment and the final objective, while the model continuously performs dozens of operations inside it. If it goes wrong or gets stuck, it examines the system logs, pinpoints the faulty code, makes the fix, and retries until the task passes.
At the training level, the greatest fear in large-model training used to be gradient explosion, which often forced engineers to get out of bed in the middle of the night and manually restore the process. Now, an agent is stationed within the training cluster. As soon as it detects an anomaly, it selects the best-matching plan from 26 historical contingency plans, automatically rolls back, corrects the data, and resumes training, with no human intervention required throughout.
At the evaluation level, reinforcement learning needs to determine whether a given round has actually made the model stronger or weaker. Moonshot AI introduced an agentic reward model specifically designed to score various "self-training recipes." High-scoring recipes remain and advance to the next round of training, while low-scoring ones are eliminated outright.
This substantial investment found validation in the leaderboards in October.
On October 14, Hillclimb released DeepRSI, an RSI-specific benchmark that tasked AI with a 48-hour autonomous research challenge. All 15 questions came from real machine-learning research tasks. Kimi K3 scored 820 points to rank third worldwide, making it the highest-ranked Chinese model and the highest-ranked open-source model, surpassing Grok 4.7 and Gemini 3.8 Flash, and trailing only Claude Fable 5.1 and GPT-6 Astra. Moreover, the cost of running K3 in this test was only $159, roughly one-quarter of the cost of Claude Fable 5.1.
However, this path of self-evolution, with its pursuit of long-horizon autonomy, soon exposed the risk of running out of control.
On October 26, less than 48 hours after the more powerful K3.1 Pro launched, the company urgently issued an apology. Because too much autonomy had been granted to agents, the frequency of calls between sub-agents spiraled out of control, rapidly draining users' quotas. The team ultimately made emergency reductions to redundant sub-agent calls and refunded 25% of the monthly quota to all subscribers.
This is in fact the shared paradox faced by every model betting on RSI: once long-horizon autonomy is unleashed, token consumption becomes difficult to keep under control.
As early as the beginning of this year, Moonshot AI CEO Yang Zhilin had clearly laid out his assessment of RSI's timeline. He believes that in 2026, large models will fully enter a phase of AI-led research: AI will automatically synthesize tasks, build environments, define reward functions, and even explore new network architectures on its own.
Along this accelerated trajectory, in 2026, Moonshot AI set out three strategic priorities:
Through technological improvements and further scaling, increase K3's effective FLOPs by at least another order of magnitude and bring its pretraining capabilities up to par with the global frontier; vertically integrate model training with the agent product experience to build a Kimi that is "truly different"; focus commercialization on agents rather than chasing sheer user scale, continually raising the ceiling of intelligence while achieving order-of-magnitude revenue growth.
A line from an internal memo by him circulated widely: "This is a noisy era, but for me, Kimi's starting point is simple: sheer curiosity, curiosity about where AGI's ceiling lies." The Kimi of 2026 has opted for the technical path that calls for a touch of obsessiveness and a touch of conviction in aesthetics.
05
Alibaba RSI: 33 Rounds of Autonomous Iteration,
AI Restructures the Training Process
In RSI, Alibaba's Qwen has focused on the path of self-evolution centered ontraining recipes, closed-loop data, and chip-model synergy, which is also the approach in China that comes closest to the direction of "models training models."
At the Yunqi Conference on September 22, Alibaba showcased the self-evolution capabilities of its core flagship model Qwen3.8-Max: it can autonomously build a training pipeline and keep it running for more than a month with no human intervention,completing 33 full rounds of frequent, effective iteration.
It uses beta-testing logs and user feedback to identify its own weaknesses, autonomously generates training data, designs controlled experiments, then has smaller models conduct multiple rounds of verification and filters out invalid samples. After this round of self-iteration, Qwen3.8-Max's score on Artificial Analysis rose sharply from 40 to 45.
Besides the closed-loop data loop on the training side,Alibaba is also advancing its exploration of chip-model synergy. On the inference side, Qwen3.8 can autonomously complete the adaptation of the next-generation Qwen3.8-Flash architecture and tune the SGLang inference framework on T-Head's brand-new GPU, directly increasing single-instance inference throughput by 96%.
At the Yunqi Conference, Wu Yongming incorporated recursive self-improvement into Alibaba's long-term technology roadmap. Before that, leaders of major domestic tech companies rarely discussed RSI directly in public strategic statements.
He believes the path toward ASI is gradually becoming clear, that recursive self-improvement is one of its main lines, and that models can continuously improve based on real-world feedback.He presented a set of projections: the total volume of machine thinking is currently less than 3% of that of humans; in the long run, it could potentially rise to 1,000 times the human level, leaving ample room for growth. He also noted that, beyond computing power, AI needs to develop perception and interaction capabilities and properly align its values. As a result, one of the team’s priorities is to connect multimodal understanding and generation.
The Qwen team has recently undergone a personnel change: Zhou Hao, a former core member of DeepMind’s Gemini self-improvement and reinforcement-learning team, has joined Alibaba.
Meanwhile, Qwen has become the foundation model for a number of external teams conducting RSI experiments. Tencent’s Terminal-RSI and BAAI’s AREX-2 are built on Qwen3.8-27B, while Chaoyan Intelligence’s Apex-35B-V1.0 uses Qwen3.6-35B-A3B.
06
ByteDance RSI: A Systematic Exploration of RSI
ByteDance Seed’s exploration of RSI is highly systematic,with its focus on research into foundational theories, evaluation, and validation mechanisms, seeking to build a rigorous set of judging criteria for the entire RSI field.
TAG750@@Volcano Engine has a dynamic model identifier, doubao-seed-evolving. It is not a fixed model, but rather the Seed team’s self-evolving R&D framework, known internally as Seed-for-Seed.
Simply put, each newly iterated version of the Seed model feeds back by generating training signals, continuously optimizing subsequent versions. This modularity allows developers to access continuously upgraded capabilities simply by integrating with the same Model ID. In the September 9 iteration, it focused on optimizing asynchronous subtask acceptance, reducing multimodal-understanding latency by 40%, and filling the gap in Office cross-application collaboration capabilities.
But behind this self-bootstrapping closed loop lies ByteDance’s most prudent scientific reflection on the evolution of RSI.
In September, Seed, together with Georgia Tech and several other leading institutions, released the Self-Developing Agents project. Together with the earlier Chain-of-Experience research, it formally formed ByteDance’s self-evolution publication portfolio. The team explicitly pointed out in its research that the vast majority of so-called RSI systems in the industry today are essentially only “semi-closed loops.”
ByteDance proposed the concept of the Golden Verifier (golden verifier) to combat hallucinations: many systems may appear to be self-cycling, but in plain terms, all the scoring rules, verification tools, and optimization directions are part of a syllabus drawn up in advance by humans. The AI simply cannot decide for itself which direction to evolve in.
To move toward a true “fully closed loop,” the team broke the challenge down into three verifiable questions:Given only a vague objective, can the model evolve autonomously; can the results of the model’s self-testing and self-evaluation be reliably transformed into tangible capabilities; and can the model independently build and continuously iterate the R&D scaffolding for agents?
The conclusions drawn from this study are relatively cautious, which is uncommon among comparable efforts by major Chinese tech companies. The experiments show that agents are highly prone to overfitting to feedback signals during self-evaluation. AI can indeed distinguish which action looks better at the moment and earns a higher score, but it finds it extremely difficult to turn this short-term judgment into a long-term, stable strategy. The high scores achieved in self-evaluation often fail to translate into lasting, reliable improvements in real-world capabilities.
Unlike other companies, Seed leans more toward foundational theory, focusing first on refining the scaffolding and evaluators. Yet, as with every approach, the overarching objective of the system is still set by humans.
07
Xiaomi RSI:
A Six-Day, Six-Night RSI Livestream
Xiaomi’s recent global livestream left many people exclaiming that it was an eye-opener. In RSI, Xiaomi did not modify its underlying framework. Instead, it started withpost-training reinforcement learningand moved the entire process online for a livestream.
The livestream lasted six days and six nights. On September 22, Xiaomi officially released and open-sourced MiMo-V2.6, positioning it as a key step in exploring the RSI path. The model scored 46 on the Artificial Analysis Intelligence Index, ranking first among open-source models, ahead of Kimi K3 with 44 and GLM-5.3 with 45.
Throughout the live reinforcement learning training run, the Pro version cost $2.62 million, while the Flash version cost $850,000. Each completed 30 iterations, generating a combined total of 750,000 training trajectories.
In terms of training configuration, the model natively supports a context window of one million tokens, samples 1,568 examples per round, and consumes 3.5 to 3.7 billion tokens per training round. To prevent expert-load imbalance under the MoE architecture, the team kept the routing module fixed. It also built a multilayer validation mechanism to address the risk of reward hacking, relying on reward design, adversarial testing, anomaly monitoring, and cross-checking by multiple verifiers to reduce the model’s attempts to game the scoring system.
Luo Fuli wrote in the livestream post:“Reinforcement learning is one of the more scalable directions on the path toward models improving themselves… We want to test this openly and see how far this path can go.”
Xiaomi made the training costs, trajectory data, and iteration process fully public. It wants to validate to the industry that, without redesigning the underlying software and hardware, open-source models can achieve meaningful self-iteration simply by committing substantial compute resources to reinforcement learning.
After taking stock of this boisterous wave of Chinese RSI, it is hard to avoid the impression of “a lot of thunder but little rain.”Not a single one is actually modifying the weights themselves; even the most aggressive players only dare to operate within the "minimal closed loop".Most players are crowding to work ontraining data, Harness, inference systems, sandbox environments, and post-training RL.
Is it because they aren't smart enough?
Exactly the opposite. This reflects the chasm between RSI's ultimate vision and engineering reality. In Tang Jie's words, this generation of RSI is still at the stage of "AI helping AI write code and AI helping AI build systems." As for the goal of enabling models to independently design the next generation of models, no one has truly reached the threshold; rashly allowing AI to modify its own weights would only plunge it into a bottomless pit of hallucinations and a loss of control over the system.
Xiaomi's six-day, six-night livestream publicly streamed the complete ledger of reinforcement learning; Zhipu's production cluster with 100,000 accelerator cards; DeepSeek's 3 million sandboxes... This appears not to touch the core of RSI. In reality, these are the stepping stones laid one after another by China's large-model vendors on the road to AGI.
When these scattered closed loops are being rapidly rolled out and the engineering closed loops that do not touch the weights are brought online one after another, the day they finally converge into a torrent will mark the qualitative transformation of Chinese foundation models.
Reference link: alphaxiv.org/abs/2609.11873
All aboard, Leiphone!(WeChat Official Account: Leiphone)Take you through the highlights of the world's top AI conferences
with exclusive access to:
Expert presentation slides
Full conference reports
Interpretations of popular papers
Interviews with rising academic stars
Scan the QR code above
or click「Read the Original Article」to follow the dedicated section.
Original article from Leiphone. Unauthorized reproduction is prohibited. See Reposting Notice for details.
