Leiphone.com Reviews the Current State of Domestic RSI: How Far Have Zhipu, DeepSeek, ModelBest and Other Vendors Each Gotten?

The real level of domestic RSI: no one dares to actually touch the model weights, but everyone is modifying "the machine that builds AI."

English text is being prepared.

English text is being prepared.

RSI, that is, AI building AI, has become an internal report card that model vendors have been eager to show off in 2026, and is also one of the hottest directions in China's primary market. As long as a BP carries the words "Self-Training," the valuation goes up by at least 20%.

Most systems at the world's leading giants are in the transition stage from L1 to L2; L3 and L4 have scattered prototypes in specific vertical domains (such as software engineering and scientific discovery), while a true L5 closed loop exists only in a very small number of restricted laboratory prototypes.

If we use this ruler to measure the RSI experimental results that domestic model vendors showcased in September and October, including training pipelines, inference systems, GPU operators, and Agent scaffolding; but at present no vendor dares to use AI to truly rewrite the underlying model weights. That is to say, everyone is currently still far from the legendary "AI inventing the next generation of AI by itself (L5)"—they have not even touched the threshold.

So in this RSI boom, where exactly are China's major giants and star startups stuck in the ecosystem? What are they each having AI improve, and how far have they gotten? Below we have sorted out the latest RSI situation of domestic model vendors.

Zhipu is the first domestic case to truly run RSI in a production environment. Its first step in RSI did not touch the model weights, but landed first on the dirtiest and most exhausting underlying inference system; after getting it running, Zhipu then set its sights on the Agent Harness. The former cuts costs, the latter boosts efficiency—all solid engineering work.

Let's start with the inference infrastructure. For a new model to actually be usable by users, there is a mountain of dirty and exhausting work in between: how to squeeze out the chip's compute power, how to split huge tasks across tens of thousands of cards to run simultaneously, how to efficiently manage the massive amounts of temporary data generated during conversations, and so on. In the past, these things were tuned line by line by Infra engineers.

This time, Zhipu directly let the "Infra Agent" driven by GLM-5.3 take over all of this. On September 17, Tang Jie posted a long article on X, "Toward Recursive Self-Improvement," stating that the Infra Agent driven by GLM-5.3, on a cluster composed of 100,000 domestic chips, built the production-grade inference service for GLM-5.3-Flash from scratch in just two weeks.

While handling all online traffic, this system pushed end-to-end throughput to 3.2 times the initial baseline, and hardware utilization and per-token cost directly caught up with the level of mainstream Nvidia GPUs. Supported by this self-built inference stack, Zhipu's anonymously launched GLM-5.3-Flash consumed more than 62 trillion tokens in 6 days.

And these proven optimization solutions will be abstracted into "optimization skeletons" and stored in an asset library, to be directly reused the next time a similar bottleneck is encountered. In Tang Jie's words: "Zhipu has already achieved the smallest closed loop: the model optimizes the system, and the system serves the model."

On the day of the news, Zhipu's Hong Kong stock rose 8.29%, and the market directly paid for the results of "Zhipu RSI."

Soon after, in late October, the Tsinghua and Zhipu team updated the paper "HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs," with Tang Jie personally listed as an author, targeting the Agent Harness.

As the scheduling framework surrounding the large model, the Harness manages prompt design, tool invocation, allocation of compute budgets, and task routing. In the past, this intricate set of control parameters could only be manually tuned by engineers. If AI were allowed to modify it itself, it would be especially prone to "peeking at the standard answers" and then cheating, producing something that scores high but performs poorly.

HarnessPro's breakthrough lies in the fact that it does not touch the standard answers at all. The system dispatches multiple dedicated Agents to collaborate: a diagnosis Agent is responsible for finding bottlenecks, a test generation Agent synthesizes its own test cases for stress testing, and a planning and implementation Agent modifies the framework code according to different orientations such as saving money or performance; at the same time, "accuracy" and "execution cost" are run together as two metrics.

As a result, under the premise of zero contact with answers, the pass rate on highly difficult code tasks did not fall but rose, and the cost per task was directly cut by 53%.

This series of actions precisely confirms the industry-famous saying of 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 that brilliant, as long as there is an object that can be improved, a verifiable judge, and a sustainable feedback loop, and these three things are meshed into a closed loop, performance can be pushed to the limit.

Zhipu's ambition is clearly not limited to this. At the semi-annual 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 about 60% of the net proceeds of the newly raised $5 billion into this self-evolution system.

In a speech in Singapore in early October, he also laid out three paths for GLM-5.4 / 5.5: scaling pretraining data to several trillion tokens, which means the parameter count is expected to surge from the current more than 740 billion to at least 1 trillion; strengthening RL post-training by deepening SFT, RLHF, and RLVR; and continuously extending long-horizon reasoning and agent capabilities.

He stressed that AI basically no longer needs to compete on dialogue or coding, and that the next generation will take two paths: RSI, autonomous iterative upgrading, and longer-cycle, more autonomous Agents.

But he did not shy away from the biggest difficulty either. "It's not continuing to scale the model," he said, "but enabling the model to reliably judge when to stop, when to correct errors, and how to evaluate its own training results."

DeepSeek's RSI is somewhat different from others: what it transforms first is the Agent's training ground—the sandbox.

RSI is essentially high-density trial and error. After one generation of Agent training is completed, engineers still have to build the next round of training environments from scratch. If this bottleneck is not solved, no matter how smart the model is, it cannot get off the ground.

To break this deadlock, on September 19, DeepSeek, together with Tsinghua University, launched the automated sandbox cluster DSec. It is a veritable "AI testing-ground factory": it has a massive base of 30,000 CPU cores and 250TB of memory, and can produce about 3 million sandboxes per day. From lightweight function calls to full virtual machines, four tiers of isolated environments respond within seconds and are available on demand.

Here, for the first time, AI has taken over its own pace of evolution, no longer relying on manually built training grounds, but forming a systemic loop in which "Agents automatically build environments—environments nurture stronger Agents—stronger Agents feed back into better environments," and all these built environments can be reused as assets.

At the same time, DeepSeek is also improving Harness.

DeepSeek Harness, open-sourced in August, focuses on "everything is a plugin," modularizing the agent's tools, logs, and even the core runtime framework. The benefit is that it enables the Agent to "upgrade while running," and if something is changed incorrectly, it can be restored with one click.

To provide solid technical support for this "plug and play at any time," DeepSeek, together with Peking University, elaborated on the underlying Cordis mechanism in the paper "A Programming Paradigm for Spatiotemporal Composability." This mechanism specifically tackles two industry deadlocks: first, if an error occurs, how to roll back to the original state; second, if a plugin is removed, how to automatically adjust dependencies without affecting other modules.

This design has been running stably for four years on the open-source chatbot framework Koishi, successfully supporting more than 4,000 community plugins.

On October 21, DeepSeek launched DeepEvolve, a self-evolving agent capable of "predicting the future," which is also a key step in RSI (recursive self-improvement): teaching AI how to self-evolve in the real world of information pollution.

Its core is to separate "information search" from "result prediction." By precisely distilling reusable prediction skills from limited real-time trajectories, even if external information is chaotic, AI can filter out interference, extract true knowledge by reviewing experience, and achieve closed-loop algorithm iteration.

Every move Liang Wenfeng makes may seem scattered, but in fact they all point to the same endgame.

From the MoE architecture and MLA attention mechanism to the DSA sparse algorithm, and now to the DSec sandbox and the self-reconfigurable Harness, all of DeepSeek's innovations are doing one thing: lowering the compute threshold for AGI and RSI.

03

ModelBest RSI: An automated factory for "building AI"

Generally, when adapting AI to chips, especially domestic chips, human engineers need to do a large amount of low-level operator tuning, communication operator rewriting, and software-hardware co-tuning. This is extremely experience-dependent drudgery that takes months and directly determines whether chip compute can be fully unleashed. ModelBest hands the development of this low-level adaptation code to AI for iterative completion. The performance of the code produced by AI can reach up to 1.6 to 5 times that of versions handwritten by top engineers, and the software-hardware adaptation cycle that originally took months has been compressed to days.

In May this year, ModelBest internally launched the "Forge Engineering" project. Li Yuxuan, chief scientist of ModelBest, summarized the project's core idea as: "Give more environment, give fewer instructions." The entire project advances step by step along the underlying software chain for building AI.

Step one: develop the training software ForgeTrain. On September 24, ModelBest announced the results: AI independently wrote the large-model pretraining framework ForgeTrain from scratch, taking only 8 hours to catch up to the benchmark level of Nvidia's mature framework Megatron-LM, and completing a performance overtake within 2.5 days. If this work were handed to humans, it would require at least 3 to 5 top engineers and more than 6 months.

Step two: develop the underlying operators, ForgeKernel. On October 27, ModelBest announced its latest progress: AI has moved further down to the very bottom layer of chips, to rewrite the "last mile" connecting models and hardware. This self-iterating ForgeKernel operator system successively took first place on the official operator leaderboards of Huawei and Nvidia. In tests on the Ascend 950 series, its benchmark speedup reached 1.6 to 5 times that of the runner-up, equivalent to effectively raising compute utilization without any increase in hardware investment.

From training frameworks and underlying operators to scientific computing, AI is gradually taking over this kind of low-level code development work. At the same time, human engineers are shifting to setting standards and evaluation metrics (Harness).

It is worth noting that ModelBest was also one of the deep participants in the September report "The Last AI Built by Humans." If ModelBest's project is placed on this roadmap, it is currently at the stage where AI is deeply involved in the engineering implementation of "AI making AI," but has not yet reached the top-level decision-making stage of "deciding for itself what to train and why to train it."

Moonshot AI's RSI takes the "internal strength" route. It pours a large amount of compute into large-scale reinforcement learning, going all out to improve AI's long-horizon autonomous capabilities.

In the post-training stage, Moonshot AI uses KDA hybrid linear attention and the MoonEP training infrastructure to let the model iterate repeatedly in real task environments such as human body simulation, finance, and 3D modeling.

On July 16, Moonshot AI released and open-sourced Kimi K3, with 2.8 trillion total parameters. To support this kind of high-intensity Agent training, the team built a sandbox system called AgentEnv; during training and evaluation, it cumulatively created more than 51 million sandboxes and used more than 1.5 million images. But under extremely high-pressure stress testing, traditional containers frequently suffered kernel crashes and deadlocks, causing training to be interrupted often. The team ultimately switched to Firecracker microVMs as the technical support.

If this system is broken down, Moonshot AI actually automated three levels.

At the task level, a human only provides a sandbox environment and the final goal, and the model continuously executes dozens of steps of operations inside it. If it goes wrong or gets stuck, it reads the system logs itself, locates the erroneous code, fixes it, and reruns it until the task passes.

At the training level, large-model training used to fear gradient explosions most, and engineers often had to get up in the middle of the night to recover manually. Now an agent is stationed in the training cluster; once an anomaly is detected, it picks the best-matching set from 26 historical contingency plans, automatically rolls back, corrects the data, and continues training, with no human intervention throughout.

At the evaluation level, reinforcement learning needs to know whether this round actually got stronger or weaker. Moonshot AI introduced an agent reward model specifically to score various "self-training recipes"; high-scoring recipes are kept for the next round of training, while low-scoring ones are eliminated directly.

This heavy investment was certified by a leaderboard in October.

On October 14, Hillclimb released DeepRSI, a dedicated RSI benchmark that gives AI a 48-hour autonomous scientific research challenge, with all 15 problems drawn from real machine learning research tasks. Kimi K3 ranked third globally with 820 points, the highest-ranked Chinese model and open-source model, beating Grok 4.7 and Gemini 3.8 Flash and trailing only Claude Fable 5.1 and GPT-6 Astra. Moreover, in this test K3's operating cost was only $159, about one quarter that of Claude Fable 5.1.

But this self-evolution route pursuing long-horizon autonomous capabilities soon exposed the risk of losing control.

On October 26, less than 48 hours after the more powerful K3.1 Pro went online, the company urgently issued an apology notice. Because too much autonomy had been granted to agents, the frequency of calls between sub-agents spiraled out of control, and user quotas were heavily wasted. The team ultimately urgently trimmed redundant sub-agent calls and refunded 25% of the monthly quota to all subscribers.

This is actually a contradiction faced by all models betting on RSI: once long-horizon autonomous capabilities are unleashed, token consumption becomes very hard to control.

As early as the beginning of this year, Moonshot AI CEO Yang Zhilin clearly laid out his judgment on the pace of RSI. He believes that in 2026 large models will fully enter the stage of AI-led research: AI automatically synthesizing tasks, building environments, defining reward functions, and even exploring new network architectures on its own.

Following this acceleration route, in 2026 Moonshot AI has defined three strategic mainlines:

Through technical improvements and further scaling, raise K3's equivalent FLOPs by at least another order of magnitude, catching up to the international frontier in pretraining; vertically integrate model training with the Agent product experience to create a "different" Kimi; focus commercialization on Agents, not pursuing absolute user scale, continuously raising the intelligence ceiling while achieving order-of-magnitude revenue growth.

In an internal letter, he had a widely circulated line: "This is a noisy era, but for me, Kimi's starting point is very simple—pure curiosity, curiosity about where the upper limit of AGI lies." Kimi in 2026 chose that technical route that requires a bit of paranoia and a bit of aesthetic persistence.

On RSI, Alibaba Qwen focuses on the self-evolution path of training recipes, data closed loops, and chip-model co-design, which is also the route in China closest to the direction of "models training models."

At the Yunqi Conference on September 22, Alibaba demonstrated the self-evolution strength of its core flagship model Qwen3.8-Max: it could, with absolutely no human intervention, autonomously build a training pipeline and keep it running continuously for more than a month, completing a full 33 rounds of high-frequency effective iterations.

It locates its own weaknesses based on internal test logs and user feedback, autonomously generates training data, designs controlled experiments, and then hands them to smaller models for multiple rounds of verification to filter out invalid samples. After this round of self-iteration, Qwen3.8-Max's score on Artificial Analysis rose sharply from 40 to 45.

Beyond the data closed loop on the training side, Alibaba is also advancing exploration of chip-model co-design. On the inference side, Qwen3.8 can, on T-Head's all-new GPU, autonomously complete architecture adaptation for the next-generation Qwen3.8-Flash and tuning of the SGLang inference framework, directly raising single-instance inference throughput by 96%.

At the Yunqi Conference, Wu Yongming incorporated recursive self-improvement into Alibaba's long-term technology roadmap. Before this, top executives of major Chinese companies rarely discussed RSI directly in public strategic statements.

He believes the path to ASI has gradually become clear, and recursive self-improvement is one of its mainlines, with models able to continuously optimize based on real feedback. He offered a set of estimates: currently the total amount of machine thinking is less than 3% of humanity's, with the long-term potential to rise to 1,000 times humanity's—a huge room for growth. He also mentioned that beyond compute, AI must also add perceptual interaction capabilities and do a good job of value alignment, so one of the team's priorities is connecting multimodal understanding and generation.

The Qwen team recently saw a personnel change: Zhou Hao, formerly a core member of DeepMind Gemini's self-improvement and reinforcement learning direction, joined Alibaba.

On the other hand, Qwen has become the base model for RSI experiments by many external teams. Tencent Terminal-RSI and BAAI AREX-2 are built on Qwen3.8-27B, while Chaoyan Intelligence Apex-35B-V1.0 uses Qwen3.6-35B-A3B.

06

ByteDance RSI: Systematic RSI Exploration

ByteDance Seed's exploration of RSI is highly systematic, with its focus on research into underlying theory, evaluation, and verification mechanisms, attempting to build a rigorous set of referee standards for the entire RSI track.

Volcano Engine has a dynamic model identifier, doubao-seed-evolving. It is not a fixed, unchanging model but the Seed team's self-evolution R&D framework, internally called Seed-for-Seed.

Simply put, the new versions of the Seed model produced by iteration will in turn generate training signals to continuously optimize subsequent versions. This modularity lets developers access continuously upgraded capabilities simply by connecting to the same Model ID. In the September 9 iteration, it focused on optimizing asynchronous subtask acceptance, reduced multimodal understanding time by 40%, and filled in Office cross-application collaboration capabilities.

But behind this bootstrapping closed loop is the ByteDance team's most cautious scientific reflection on RSI evolution.

In September, Seed, together with Georgia Tech and several other top institutions, released the Self-Developing Agents project, which, along with the earlier Chain-of-Experience research, formally formed ByteDance's self-evolution paper matrix. The team explicitly states in the research: the vast majority of so-called RSI systems in the industry today are essentially only "semi-closed loops."

ByteDance proposed the concept of a Golden Verifier to break the illusion: many systems now appear to be self-looping, but plainly put, the scoring rules, verification tools, and optimization directions are all an exam syllabus drawn up in advance by humans. AI simply cannot decide for itself in which direction to evolve.

To move toward a true "full closed loop," the team broke it down into three verifiable questions: given only a vague goal, can the model evolve autonomously; can the results of the model's self-testing and self-evaluation stably turn into real, solid capabilities; and can the model build its own R&D scaffolding for agents and continuously iterate on it.

The conclusions of this research are relatively cautious, which is fairly rare among similar explorations by major Chinese companies. Experiments show that agents easily overfit to feedback signals during self-evaluation. AI can indeed tell which action looks better right now and can score high, but it is extremely difficult for it to solidify such short-term judgments into long-term, stable strategies. The high scores obtained in self-evaluation often cannot be converted into long-term, reliable capability improvements in the real world.

Unlike other vendors, Seed leans more toward underlying theory, polishing the scaffolding and evaluators first. As with all approaches, the top-level goals of this system are still set by humans.

Xiaomi's recent livestream, broadcast across the entire internet, left many people saying it was an eye-opener. On the RSI front, Xiaomi did not touch the underlying framework; instead, it started from the post-training reinforcement learning stage and streamed the entire process online.

The livestream lasted 6 days and 6 nights. On September 22, Xiaomi officially released and open-sourced MiMo-V2.6, which the company positions as a key step in exploring the RSI route. The model scored 46 on the Artificial Analysis Intelligence Index, ranking first among open-source models, above Kimi K3's 44 and GLM-5.3's 45.

Throughout the real-time reinforcement learning training, the Pro version invested $2.62 million and the Flash version invested $850,000, each completing 30 iterations and generating a cumulative 750,000 training trajectories.

In terms of training configuration, the model natively supports a million-token context window, samples 1,568 examples per round, and consumes 3.5 billion to 3.7 billion tokens per training round. To avoid expert load imbalance under the MoE architecture, the team locked the routing module; at the same time, to address reward hacking risks, it built a multi-layer verification mechanism that relies on reward scheme design, adversarial testing, anomaly monitoring, and cross-checking by multiple verifiers to reduce opportunistic score gaming by the model.

Luo Fuli wrote in the livestream post: "Reinforcement learning is one of the more scalable directions on the path of model self-improvement... How far we can go on this path, we want to try publicly once."

Xiaomi made the training costs, trajectory data, and iteration process all public. It wants to verify to the industry that without modifying the underlying software and hardware, relying on saturated compute investment in reinforcement learning, open-source models can also achieve considerable self-iteration.

Having taken stock of this vigorous wave of domestic RSI, one cannot help but feel it is "much thunder but little rain." Not a single company is really changing the weights themselves; even the most aggressive players only dare to circle within a "minimal closed loop." What everyone crowds around to change is mostly training data, Harness, inference systems, sandbox environments, and post-training RL.

Is it because they are not smart enough?

Quite the opposite. What this reflects is the gap between the ultimate vision of RSI and engineering reality. In Tang Jie's words, this generation of RSI is still at the stage of "AI helps AI write code, AI helps AI build systems"; as for the goal of letting models independently design the next generation of models, no one has truly reached the threshold yet, and rashly letting AI modify its own weights would only lead to a bottomless pit of hallucinations and system loss of control.

Xiaomi's 6-day, 6-night public livestream of the complete reinforcement learning ledger, Zhipu's 100,000-card production cluster, DeepSeek's 3 million sandboxes... This may look like it has not touched the soul of RSI, but in reality it is the stepping stones that Chinese large-model vendors are laying one by one on the road to AGI.

When these scattered closed loops are being rolled out at an accelerating pace, and the engineering closed loops that "do not touch the weights" are connected segment by segment, the day they finally converge into a torrent will be the moment Chinese large models undergo a qualitative change.

Reference link: alphaxiv.org/abs/2609.11873