WeChat's WeLM targets an agent-driven ecosystem with 80B and 617B models

What’s most noteworthy about WeLM isn’t its 617 billion total parameters, but that WeChat has clearly focused the model’s goals on agents, tool invocation, and execution within the WeChat ecosystem: the 80B version has entered the “small” category, while the 617B version directly targets more complex mini-program development and tool generation.

Article author, source: 0x9999in1, ME News



• What’s most noteworthy about WeLM isn’t its 617 billion total parameters, but that WeChat has clearly focused the model’s goals on agents, tool invocation, and execution within the WeChat ecosystem: the 80B version has entered the “small” category, while the 617B version directly targets more complex mini-program development and tool generation.
• Total parameters: 80B (80 billion), with approximately 3 billion activated per token; 617B (617 billion) total parameters, with approximately 23 billion activated. What truly determines commercial significance is not how massive the total parameters are, but how much computational cost each task requires.
• The Hidden Decoding experiment revealed in July showed that the WeLM team is exploring an alternative scaling path: rather than merely increasing parameters or extending explicit reasoning chains, they are introducing implicit computation within the model. The results are promising, but the costs are equally real, particularly with noticeable throughput degradation in long-context and large-batch scenarios.
• WeLM-80B improved from 44.9 to 58.4 on Terminal-Bench 2, and ARC-AGI-2 increased from 6.9 to 11.6. Such improvements in agent performance and long-range task metrics better explain why WeChat is interested in this approach than simply ranking higher on knowledge-based QA leaderboards.
• In the second quarter of 2026, Tencent's capital expenditures reached RMB 52.78 billion, a 176% year-over-year increase; research and development expenses amounted to approximately RMB 27.2 billion, up 35% year-over-year. WeLM was developed within a context of massive AI infrastructure investment, not in a lab with "free compute power."
• My assessment is that WeLM is not Tencent’s sole solution for unifying all its AI capabilities; rather, it resembles an AI engine specifically built by WeChat for its super app. If “WeChat Mini” ultimately succeeds, WeChat’s core competitive approach may gradually shift from “users actively seeking services” to “AI organizing and invoking services on behalf of users.”

On August 13, 2026, the WeChat team further disclosed two configuration options for WeLM.

WeLM-80B, with a total of 80 billion parameters and approximately 3 billion parameters activated per inference, has been integrated into WeChat's native AI Agent "Xiao Wei," handling tasks such as chatting, searching, and accessing WeChat's native capabilities and mini-program services.

The larger WeLM-617B is still under development. With a total of 617 billion parameters and approximately 23 billion activated, it is no longer focused solely on “answering questions,” but rather on more complex WeChat ecosystem tasks, including intelligent development of mini-programs and tools for micro-businesses.

Seeing the number 617 billion, people might instinctively wonder: Is WeChat also jumping into the race for massive models?

But this may precisely miss the point.

617B is certainly large, but one of the most important economic parameters for MoE models is never "how many parameters are stored in total," but rather how many parameters are actually activated during the processing of each token. With 617 billion total parameters and 23 billion activated parameters, the activation ratio is under 4%; for 80B, the corresponding 3B activated parameters also follow a very sparse approach.

Why does WeChat care so much about this?

Because the large models in the lab and the large models in WeChat are not the same kind of business.

As of the end of June 2026, the combined monthly active accounts of WeChat and WeChat reached 1.439 billion, a 2% year-over-year increase. When AI capabilities are integrated into a product of this scale, even a slight increase in cost per request, multiplied by hundreds of millions of users and high-frequency usage, results in massive server, GPU, power, and depreciation expenses. Tencent’s Q2 financial report, released on the evening of August 12, clearly articulated this pressure: capital expenditures for the second quarter reached RMB 52.78 billion, a year-over-year increase of 176%; operating capital expenditures reached RMB 51.8 billion, up 190% year-over-year; research and development expenses amounted to approximately RMB 27.2 billion, up 35% year-over-year; free cash flow declined to negative RMB 13.8 billion.

So the issue before WeChat has never been just: Is the model smart enough?

Here’s another more realistic question: After becoming smart, can you afford to support yourself?

This is also the first key to understanding WeLM. WeChat did not write “resource efficiency” as a catchy slogan in the model’s description—it faced the nearly brutal cost multiplication that comes with launching a product for a billion-plus user base.

If you look only at model size, another question may easily arise.

Since 617B is clearly stronger, why not just put the largest model into Xiao Wei?

The answer is already hidden in the tasks WeChat has assigned to the two models.

handles chat, search, native WeChat functions, and小程序 service calls. The primary requirements for these tasks are not necessarily peak intelligence, but rather fast response, low cost, stable invocation, and accurate understanding of WeChat's internal objects and operations.

You ask it to find information, invoke a service, or enable a feature—the thing users truly can't tolerate isn't the model not knowing a math olympiad problem, but waiting over ten seconds only for it to click the wrong button.

617B targets another layer of tasks: more complex understanding, reasoning, code generation, tool building, and even mini-program development. The longer the task and the more steps involved, the greater the cumulative loss from a single incorrect judgment—making it worthwhile to invest additional computational cost for stronger reasoning capabilities.

This is actually a very typical agent economics model: simple tasks aren’t worth deploying the most expensive brainpower—only complex tasks require upgraded computational resources.

This also means that a truly mature WeChat AI in the future is unlikely to rely on a single model to handle everything, but rather a system that orchestrates multiple models, tools, and computational budgets. Whether WeChat will ultimately adopt such dynamic routing remains unconfirmed by any public information; however, the fact that the 80B and 617B models are assigned distinctly different product roles already indicates that WeChat does not intend to solve all problems with a simple “brute force” approach.

This judgment is important.

Over the past two years, the most talked-about stage in the large model industry has been rankings: who has more parameters, who scores higher on math tests, and who has better code benchmarks. But as agents enter real-world products, the evaluation system is changing.

Users won't ask every day: "What is your GPQA Diamond?"

The user will only ask: "I said one thing—can you or can you not get the job done?"

WeChat happens to possess the richest collection of “things that can be done” in all of China: messaging, content, official accounts, mini-programs, payments, shopping, local services, and various capabilities provided by developers. What truly sets WeLM apart is not that it has simply created another model, but that it was born sitting right next to a pile of tools.

WeLM-80B and 617B are not names that suddenly appeared today.

The paper submitted on July 9, "Hidden Decoding at Scale: Latent Computation Scaling for Large Language Models," already presents two WeLM architectures: the 80B model with 80 billion total parameters and 3 billion activated parameters, and the 617B model with 617 billion total parameters and 23 billion activated parameters, corresponding to 49 and 94 layers respectively.

What's truly interesting is the question the paper attempts to answer.

How can large models continue to become smarter?

There are two most familiar methods.

First, make the model larger.

Second, let the model "think longer" by using more inference tokens and a longer explicit reasoning process, trading computation for capability.

Hidden Decoding attempts to insert a third path: expanding a single token into multiple internal computation streams without further expanding the Transformer backbone, allowing the model to perform several rounds of invisible internal computations before outputting the next token.

You can roughly think of it this way: the traditional model receives a question, passes it through the neural network once, and then starts writing; Hidden Decoding, however, wants it to draft multiple times internally before actually writing anything down.

This is not a metaphysical metaphor. The WeLM team used four-way Hidden Decoding, also known as HD4, in their experiments.

The results are well worth seeing.

Under the same early supervised fine-tuning conditions without reinforcement learning, WeLM-HD4-80B achieved a 13.5 percentage point increase in solve rate on Terminal-Bench 2, rising from 44.9 to 58.4; ARC-AGI-2 improved from 6.9 to 11.6; and LiveCodeBench v6 increased from 83.7 to 85.1. Eight out of ten extended tests showed improvements.

Why is the number Terminal-Bench 2 particularly noteworthy?

Because it doesn't require the model to simply sit and answer a multiple-choice question, but instead asks the model to operate and perform tasks within a container environment, with the final validation determining whether the task is completed. In other words, it is significantly closer to how an Agent actually works.

The results of 617B were also impressive. In all nine challenging benchmarks provided in the paper, the HD4 version outperformed the standard 617B version. For example, GPQA Diamond increased from 89.1 to 91.2, HLE from 33.6 to 35.4, and SciCode from 51.4 to 52.1.

The paper also uses Kimi K2.6, with 1 trillion total parameters and 32 billion activated parameters, as an external reference. WeLM-HD4-617B outperforms it in five of nine metrics and underperforms in four.

How should this set of data be described?

Not "completely overwhelming".

More accurately, WeLM-HD4-617B has already demonstrated competitive performance with state-of-the-art MoE models at significantly lower activation parameter counts.

This wording is less sensational, but more important. For WeChat, every inference could ultimately land on the cost sheet.

The moment an AI technology is most likely to be misunderstood is when it showcases improved capabilities without mentioning the bill.

Hidden Decoding also has a bill.

Theoretically, expanding the sequence by four times could cause attention costs to skyrocket if all attention layers perform full computations. Therefore, the WeLM team designed Stream-Factorized Attention, which allows most layers to compute only within a single stream, while only a subset of layers perform cross-stream information exchange.

The final measured training cost did not inflate to the 16-fold increase that could occur under full attention: in the 80B experiment, a 4-fold increase in effective sequence length resulted in a training time per batch of approximately 5.1 times the baseline; for 617B, it was about 4.4 times.

Much better already, but still not a free lunch.

The inference side better illustrates the issue.

When tested on eight H20 GPUs with WeLM-HD4-80B in an interactive scenario with a batch size of 1, the model maintains a decoding throughput of approximately 83% to 88% of a standard 80B model. This is critical, as chat-based products closely resemble workloads with low batch sizes and a focus on response time.

However, as the context lengthens and the batch size increases, the cost rapidly becomes apparent: with a 32K input and a batch size of 32, throughput drops to just 27% of the baseline.

So, if someone interprets Hidden Decoding as “gaining four times the thinking ability for free,” that’s definitely going too far.

What it truly proposes is another way of resource exchange: allowing the model to expend more internal computation in exchange for enhanced capabilities without further increasing the backbone parameters, while ensuring that this additional computation remains compatible with the parallel training architecture of modern large models.

This is an engineering choice, not magic.

One more important point: Without official clarification, the WeLM-HD4 model described in the paper cannot be simply equated with the WeLM model announced by WeChat today and used in product introductions. The paper clearly distinguishes between the standard WeLM and the HD4 version with Hidden Decoding. Current public information confirms only that the 80B model is powering Weixiao, along with two baseline architectures and product positions; it cannot be concluded that all online Weixiao requests are running on HD4.

This boundary must be kept.

The biggest mistake in technical analysis is seeing two puzzle pieces that look similar and assuming you know where the final piece belongs.

Even now, I still believe the easiest mistake when discussing WeLM is imagining "WeLM" as just another chatbot inside WeChat.

That underestimates WeChat.

The chat box is just the most natural entry point for AI, but not necessarily the end.

Tencent's earnings report defines micro-businesses as "Agentic AI within WeChat," and explicitly states that WeLM is a model customized for user privacy, WeChat scenarios, and inference efficiency. Meanwhile, WeChat has today further specified the 80B task: invoking native WeChat functions and accessing mini-program services.

Putting these two pieces of information together makes the direction quite clear.

In the search era, users look for pages.

In the app era, users seek functionality.

The most unsettling yet exciting change in the Agent era is that users may not even want to look for features anymore.

To order something, check a service, invoke a mini-program, organize content, or even generate a temporary tool in the future, the user simply expresses their intent, and the Agent determines which services to invoke.

At this point, WeChat’s accumulated mini-programs and native services over the past decade are no longer just a row of icons requiring users to open them manually—they have the potential to become “capability interfaces” that AI can orchestrate.

This is where WeLM truly raises concern.

The model itself is not the entire moat; it becomes the moat when the model can access what it needs.

A model that leads by 5 points on an open benchmark but can only answer questions; another slightly weaker one that can safely and reliably handle payments, content, social functions, and services provided by millions of developers—which one ultimately has greater commercial value?

The answer may not belong to the top-ranked position.

Tencent's own financial figures already demonstrate this logic. In Q2, marketing services revenue reached RMB 43.6 billion, a 22% year-over-year growth, with the company explicitly attributing the growth to upgrades in AI advertising recommendation models, AIM+ intelligent placement, and the closed-loop marketing capabilities within the WeChat ecosystem. In other words, before consumer agents achieve large-scale commercialization, AI has already translated into revenue growth within Tencent's existing businesses.

This is also the biggest difference between large internet platforms and pure model companies.

It doesn't need to wait for AI itself to start charging subscriptions to prove its value.

WeLM still faces three very practical challenges.

The first is reliability.

The chatbot may give one incorrect answer, and the user can ask once more; but if an agent calls the wrong service or executes a step incorrectly, the cost is much higher. As the model moves from “speaking” to “doing,” the tolerance for error doesn’t just decrease slightly—it fundamentally changes the entire product’s responsibility.

The second is permissions.

WeChat possesses an extremely rich network of relationships, content, services, and transaction scenarios. The deeper the AI can access, the more useful it becomes; however, the deeper the access, the more challenging it becomes to design privacy, security, and authorization systems. Tencent's strong emphasis on customizing WeLM for user privacy clearly indicates that this is not a peripheral issue, but a fundamental condition for whether small models can truly scale.

The third is still cost.

Tencent's capital expenditure of RMB 52.78 billion in the second quarter has sent a clear signal to the entire market: the so-called "battle for entry points" in the AI era is first and foremost a capital-intensive war.

More interestingly, Tencent reported a Non-IFRS operating profit of RMB 75.6 billion; however, if the impact of new AI products such as Hy, Yuanbao, CodeBuddy, WorkBuddy, and Xiaowei is excluded, the profit reaches RMB 86.1 billion, a difference of approximately RMB 10.5 billion.

This cannot be simply explained as "AI lost RMB 10.5 billion," as this figure includes the combined impact of related revenue, costs, and expenses; but it at least makes one thing clear: Tencent is spending a substantial amount of real money—enough to significantly affect its profit growth rate—to buy a ticket to the next generation of products.

So, it's still too early to judge the success of WeLM today.

617B is still under development, and Xiao Wei remains in limited prototype testing and grayscale phase. Model evaluations can tell us it "has the capability," but they cannot prove whether users in a 1.4 billion-person product would be willing to use it daily, nor can they prove whether the Agent can consistently and stably complete real-world transactions and complex tasks.

But since today, one thing has become increasingly clear:

WeChat is not just adding an AI button to the chat interface; it is attempting to make AI a systemic layer that reorganizes services within WeChat.

80B is responsible for making this system run.

617B is trying to make it handle more difficult tasks.

Hidden Decoding is exploring whether the same brain can use new computational methods to think even more.

Behind these three things is actually the same goal: finding a balance between intelligence, cost, and execution capability sufficient to serve a billion users.

What’s truly worth waiting for is not the exact day 617B goes live, nor who leads by half a point on the next Benchmark leaderboard.

Instead, one day, when users want to do something, will they越来越少 say, "Which mini-program should I open?" and instead just say to WeChat:

I've taken care of it for you.

If that day ever comes, WeLM's greatest significance will not be as a model.

It will be the engine that propels WeChat from a "super app" to a "super agent."

1. Tencent Holdings Limited.Tencent Announces 2026 Second Quarter Results. 2026-08-12. Tencent discloses Q2 revenue, profit, capital expenditures, monthly active users for WeChat and WeChat, micro-grayscale testing, and the positioning of WeLM.
2. Liu, Aiwei, et al.Hidden Decoding at Scale: Latent Computation Scaling for Large Language Models. arXiv:2607.08186, 2026-07-09. Includes WeLM-80B, WeLM-617B architectures, Hidden Decoding evaluation, training cost, and inference throughput experiments.
3. Su, Hui, et al.WeLM: A Well-Read Pre-trained Language Model for Chinese. arXiv:2209.10372, 2022-09-21. Early public research in the WeLM series.
4. IT之家: "Tencent WeChat Announces WeLM Model: 80B Version Already Deployed on AI Assistant Xiao Wei, 617B Version Under Development," 2026-08-13. The report details the two parameter scales and product positioning revealed in WeChat team's promotional poster.
5. Reuters.Tencent Q2 revenue climbs 11% on AI-driven ad gains, but profit falls short. 2026-08-12. An independent report on Tencent's Q2 performance, AI capital investment, and advertising revenue growth.