Using Alibaba as the underlying base, Tencent Hunyuan has achieved "Small RSI"
Since the release of Hy4 Preview, its most notable highlight has been the full-link RSI, which also marks the first model of Tencent Hunyuan that the team has participated in independent R&D throughout the entire process. The full-link RSI can bring substantial performance improvement to the model.
Less than two months later, the Frontier team of Hunyuan released Terminal-RSI. This version is more interesting: it enables a model to break through the limits of its original framework via RSI, and gain capabilities it did not possess before.
It is just like you have a microwave oven, whose original function is to heat food. After adding Terminal-RSI, not only can the original heating performance of the microwave oven be optimized, you can also use it to watch TV, listen to music and play games.
Take Terminal-Bench 4.0 as an example, which is one of the most rigorous practical terminal benchmark test sets for Agents at present. The base model used in the paper did not have the capability to run this test set at all, with a score of 0% originally.
With Terminal-RSI, this base model can reach a score of 19.84%, on par with GPT-5.6 Terra and Gemini 3.8 Flash.
There is also an interesting detail: the base model used in the paper is Qwen3.8-27B, which means Tencent and Alibaba have achieved an unexpected "cooperation" on RSI.
The paper itself also admits that this set of RSI has only been tested on a 27B-parameter model, without verification on larger models. It only has three evaluations and no independent training seeds. Therefore, there is no answer yet about what will happen when it is applied to large-parameter models such as Hy4, and whether RSI can still be implemented in that scenario.
What is Terminal-RSI?
The object trained by Terminal-RSI is called a terminal Agent. It is quite different from ordinary Agents: the Hunyuan team provides it with a sandbox and a task description, and it needs to act like an engineer to check the repository, install dependencies, modify codes and run tests in the command line, and finally verify whether the state in the sandbox can pass the check.
The paper states that there are usually two major challenges to realize RSI.
The first one is extremely slow feedback. The paper points out that the average number of interaction rounds in the whole training process fluctuates between 50 and 117. The model may make a mistake in the second step, but it cannot know whether the result is correct until the very end.
The second problem is that it is impossible to distinguish right from wrong in a granular way. The paper mentions that the validator passes 9 out of 11 checks, but this result is very difficult to be reflected in the terminal judgment.
It is just like playing a game: you win only after you destroy the enemy's crystal. You get one coin for killing a minion, and many players do not push the crystal but keep killing minions for fun, getting a very high final score. But the problem is that you can never win the game as long as the crystal is not destroyed.
The core purpose of Terminal-RSI is exactly to solve these two problems.
The workflow of Terminal-RSI is as follows: Base → RSR-1 → RPR-1 → RSR-2 → RPR-2, forming two cycles of "rewrite + refinement". The base model is the aforementioned Qwen3.8-27B, and the final output is named Terminal-RSI-27B.
The first step RSR (Recursive Self-Rewrite) only updates the actor, which is the policy that actually executes the tasks.
Its working mechanism is: the model attempts tasks in various different harnesses, and only retains the successful trajectories that pass the full verification. Then the model itself acts as a planner, and distills the successful process into a runbook.
At this stage, there is a dedicated review Agent that only checks the problem description and the runbook, verifying the format, output, and whether the validator answers or standard solutions are leaked.
After passing the review, the model re-executes the task in a brand new sandbox. Only the trajectories that still pass the verification in the re-execution and have no content leakage will be retained.
When performing supervised fine-tuning finally, the runbook and review dialogues will all be erased, leaving only the action sequence.
RSR is just like clearing a new game level: you die hundreds of times, and finally you beat the level once. Then you export the recording of that successful run, replay it over and over again, and memorize every operation step by step to form muscle memory.
The next time you enter the level, your hands will move automatically, and your mind is full of the routes of that successful run, without even needing to think.
This is the essence of RSR: extract that "one successful run", write it down as notes, and practice repeatedly according to it. The scorer does not change, the referee does not change, the rules remain the same, and the only variable is you who are going through the level. You are copying your own full-mark test paper.
The second step RPR (Recursive Policy-Refine) is reinforcement learning (PPO) with both the actor and the critic participating.
The paper emphasizes that after rewriting, the actor's preferred actions change, and the historical trajectories it encounters also change. The original critic may no longer make accurate estimations. If you directly reset the critic, you will lose the already learned experience of "what kind of situation is closer to success".
However, if you inherit the critic as it is without calibration, the old estimation errors will be mixed into the advantage values.
For example, when you practice and switch to a new set of gameplay, the old recordings no longer work well. You need to invite an analyst (critic) to help you analyze what to do next to clear the level.
But here comes the problem: the database in the analyst's mind is still for your old gameplay, and it cannot understand your current new gameplay at all.
The solution of RPR is that you stay where you are first, let the analyst re-train for 4 rounds separately with your recent performance records, so that it can understand your new gameplay and new rules. After the analyst fully gets the hang of it, you start to improve your performance together with the analyst, adjusting while practicing. This is the so-called "warming up with the coach".
So what are the final results?
Qwen3.8, a small model with only 27B parameters, has only been trained for two rounds. Its score on Terminal-Bench 2.1 rises from 51.7% to 86.5%, its score on Terminal-Bench 4.0 rises from completely incapable of finishing the tasks (0%) to 19.8%, and the average score of four tests rises from 20.7% to 50.6%, more than doubling.
Hy Frontier Returns to the Mainline of Hunyuan
There are three first authors of the paper, namely two interns of Hy Frontier, Yang Junyao and Li Zongxia, and Shi Yucheng, a Tencent research scientist based in Seattle.
The three of them have collaborated before, on a paper titled *T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks*.
This paper can be regarded as the "prequel" of Terminal-RSI, which introduces how to use reinforcement learning to train a large model into an agent that can actually work in the Linux terminal.
Before Yao Shunyu joined, the main force of Tencent's basic research was AI Lab established in 2016, which was initiated and prepared by Yao Xing. In 2017, machine learning scholar Zhang Tong was invited to serve as its first head.
However, according to the report of LatePost, Zhang Tong came with high expectations, believing that Tencent's massive business and data would become research resources unavailable in the academic circle. But after he joined, he still could not get access to the required data, and the research could only be conducted behind closed doors. He left disappointedly two years later.
After that, the head of AI Lab was replaced by Zhang Zhengyou, and then replaced by Jiang Jie in June 2024.
The same situation applies to Hunyuan: when the project was launched in 2023, the project team did not even have a single GPU. Finally, 2000 GPUs were allocated from the advertising department, half for pre-training and half for post-training.
The first head of Hunyuan was Zhang Zhengyou, whose research direction is computer vision. The second head was Jiang Jie, who has long been responsible for the product technology management of Tencent's advertising platform, has long been engaged in big data platforms, and is good at advertising recommendation. Almost none of them had experience in reinforcement learning.
By mid-2025, the team shifted its focus to benchmark ranking. An interviewee told LatePost, "We still do our work as usual, but everyone clearly knows that this approach cannot last."
Everything changed until Yao Shunyu joined Hunyuan. In an interview in June, Yao Shunyu mentioned that today's AI has three pillars: Foundation, Product and Frontier. It is necessary to build a balanced "triangular organization", because "we have not done enough cutting-edge exploration today".
The old research institutions were "too far away from the model", while the model team only "focused on chasing benchmark rankings".
Yao Shunyu's solution is to move the research directly into the model department. AI Lab was merged into the large language model department, and the original researchers became the Frontier direction of the department, belonging to the same department and the same reporting line as pre-training, post-training, evaluation and Infra.
The previous practice separated research in AI Lab, products and models in other departments, which relied on coordination and communication between different teams. The new practice integrates the Frontier team and the main model team in the same system, and the hierarchy is compressed into three layers: "Yao Shunyu — Direction Lead — Researcher/Intern".
Since then, Frontier is no longer an independent laboratory separate from Hunyuan. After the basic model department was established in July, it was incorporated into the same department, and all teams report to Yao Shunyu.
Although theoretically Frontier "sets no KPI and does not assign fixed research topics", this research line of terminal Agent plus reinforcement learning is not "boundless" research.
At the end of 2025, Tencent Hunyuan began to focus on AI programming. Therefore, Yao Shunyu built a reinforcement learning infrastructure shared by the whole group, allowing all businesses to train their own models on it, and then the data will flow back to Hunyuan. What T1 and Terminal-RSI are doing is exactly the long-horizon tasks and reinforcement learning in the terminal.
Plus that Hy 4 Preview has always emphasized full-link RSI, this Terminal-RSI is exactly a supplement of the Frontier team to the main product line of Hunyuan.
Hunyuan's research and product development are no longer separated.
Hy4 Promises a Bright Future
The full-link RSI of Hy4 Preview can be subdivided into four links: data strategy, training method, evaluation system and underlying operator. As mentioned earlier, RSR of Terminal-RSI corresponds to the data strategy link, while RPR mainly optimizes the training method.
No matter how strong the base of Hy4 is, the "completely successful" trajectories in terminal long-horizon tasks are still very scarce. RSR can generate high-quality training data by itself, and meanwhile isolate benchmark pollution.
Therefore, its significance for Hy4 is to solidify the state of "occasionally getting the task right" into the state of "stably completing the task", especially in scenarios such as debugging and system management that require multi-step reasoning plus tool calling.
The real key part is RPR, which almost perfectly meets all the requirements of Hy4.
As mentioned earlier, RPR means "warming up with the coach" to adapt to new environments, new gameplay and new rules. Since Yao Shunyu joined Hunyuan, the model has been focusing on context, which means the model often needs to face various new types of tasks.
Many large models have very high scores on benchmarks, but they always fall short when actually completing real tasks.
Because score cheating itself is a kind of skill. OpenAI's report once mentioned that the stronger the model is, the better it is at "acting". It can get all the intermediate steps, preconditions and intermediate outputs correct, and brush a large number of passing checkpoints, looking as if it has really completed the task. But the final critical result is wrong.
You cannot win the game without destroying the crystal. And Yao Shunyu is a person who does not care about benchmark rankings, but only cares about whether the model can actually complete tasks.
In January 2026, Yao Shunyu commented on DeepSeek, saying: "One of the things DeepSeek does well is that they may not pay that much attention to the numbers on the benchmarks. DeepSeek may pay more attention to: first, what is the right thing to do; second, what is the thing that they can personally experience whether it is good or bad."
Weak models do not even have the ability to cheat scores, but Hy4 is a strong model. If it acts, its performance will be even more realistic, and the process scores it brushes will be even higher. Therefore, this set of rules of "success first, process scores can never beat full marks" is specifically designed to solve this problem.
This article is from the WeChat Official Account "Letter AI", written by Miao Zheng, and published by 36Kr with authorization.