Tencent (HKEX: 0700) released a preview version of a new open-source AI model aimed at software engineering, research, and financial analysis tasks on Friday, August 28, 2026, according to a post published on Hugging Face. The release, reported first by Reuters, marks the latest move in Tencent’s open-source AI model strategy and lands the company squarely in a crowded field of Chinese labs racing to give away powerful coding and reasoning models for free while charging for the infrastructure and cloud services that run them.
Reuters reported that Tencent plans to integrate the new model alongside its own developer and productivity tools, including CodeBuddy and WorkBuddy, tying the open release directly to products it already sells to enterprise customers. That detail matters more than it might look at first glance: rather than treating open-source releases as pure marketing, Tencent is using them as the backbone of paid software it ships internally and externally.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What We Know About the August 28 Tencent Hy3 Preview
According to Reuters’ report, the newly released model is a mixture-of-experts (MoE) system with roughly 770 billion total parameters, of which about 49 billion are active per request. That activation pattern is the whole point of a mixture-of-experts design: instead of running every parameter for every query, the model routes each request through a smaller subset of specialized “expert” sub-networks, which keeps inference costs down while still giving the system a very large total knowledge base to draw from.
The model was posted to Hugging Face under the repository name tencent/Hy3, with Tencent also publishing a quantized variant, Hy3-FP8, aimed at teams that want to run the model with a smaller memory footprint. Tencent has a pattern of seeding its open releases across multiple platforms simultaneously rather than relying on a single channel: earlier Hy3-family releases went out on Hugging Face, ModelScope, GitCode, and CNB at the same time, a distribution approach that reflects how Chinese AI labs try to reach both international and domestic developer audiences in one move.
Reuters’ framing of the release is notable for what it emphasizes: not just coding, but software engineering, research and financial analysis. That three-part target list signals Tencent is not positioning this purely as a code-completion tool competing with GitHub Copilot or Cursor. It is explicitly pitched at longer, more open-ended analytical work, the kind of task where a model has to hold a large amount of context, reason across multiple steps, and produce structured output that a human analyst or engineer can act on directly.
Inside the Model: Parameters, Architecture, and Why MoE Matters
The mixture-of-experts approach that Tencent has used across its Hy3 lineage is the same basic architecture family that DeepSeek, Alibaba’s Qwen team, and Zhipu AI’s GLM line have all converged on for their largest models. Under the hood, OpenRouter’s July 2026 documentation of the model put Hy3’s MoE layers at 192 experts with top-8 routing per token, a fine-grained split that lets Tencent scale total parameter count without forcing every request through the whole network. The logic is straightforward: training and running a dense model with hundreds of billions of active parameters gets expensive fast, but an MoE model can pack far more total knowledge into its weights while keeping the actual compute cost per query much lower, since only a fraction of the network activates for any given prompt.
That tradeoff explains why the reported 770-billion-parameter total with a 49-billion-parameter active count is a meaningful number rather than just a marketing figure. A model that size sits in the same general class as GLM-5.2, which Zhipu AI has put at roughly 753 billion total parameters, and considerably below DeepSeek-V4-Pro’s reported 1.6 trillion parameters. Tencent’s earlier full Hy3 release, which shipped on July 6, 2026, was smaller still: 295 billion total parameters with 21 billion active and a 256K-token context window, plus what the company described in an August 2026 update on Hugging Face as a 3.8-billion-parameter MTP layer (multi-token prediction) stacked on top of the 21 billion active parameters to speed up generation.
That July release is where Tencent’s specific benchmark claims come from. According to a report from Japanese tech outlet GIGAZINE, Tencent said the July 6 version of Hy3 recorded scores comparable to high-performance open models such as GLM-5.2 and DeepSeek-V4-Pro, and outperformed GPT-5.5 specifically on FrontierScience-Olympiad, a benchmark designed to measure scientific research reasoning rather than general chat ability. The same report noted Hy3 completed document-processing tasks using 47.4% fewer tokens than GLM-5.2, a token-efficiency claim that would translate into a meaningful cost advantage for developers running the model at scale through an API.
From Hunyuan-Large to Hy3: Tencent’s Open-Source Timeline
Friday’s release did not come out of nowhere. Tencent has been building its open-source AI portfolio under the Hunyuan brand for close to two years. Its Hunyuan-Large model, published on GitHub in October 2024, was described at the time as the largest open-source Transformer-based mixture-of-experts model in the industry, with 389 billion total parameters and 52 billion active parameters. That release established the pattern Tencent has followed since: publish the weights, publish the training scripts, and let outside developers fine-tune and deploy the model rather than gating it behind a single paid API.
The Hy3 name itself first surfaced publicly in a preview Tencent unveiled with open weights on April 23, 2026. Bloomberg reported at the time that Tencent called the preview its most powerful yet and said it represented a big advance from complex reasoning to coding, noting the model supported OpenClaw and was made available via a wide release. Tencent followed up with a press release the next day, April 24, 2026, formally open-sourcing that preview’s 295 billion total parameter, 21 billion active parameter, 256K-context specs, and Caixin Global reported that Tencent’s own Yuanbao chatbot had already migrated its primary model from DeepSeek to the Hy3 Preview by that same month. That April preview was followed by the full Hy3 launch on July 6, 2026, which shipped under the Apache 2.0 license, a notably permissive choice that places no field-of-use or geographic restrictions on how developers can use the weights, according to a summary published by AI research site DigitalApplied.
The gap between the July release (295 billion total parameters) and the figures Reuters reported for Friday’s new preview (770 billion total parameters) points to Tencent scaling up its open-source flagship considerably within a matter of weeks, rather than simply reissuing the same model. Some naming ambiguity remains in the public record: multiple outlets and the Hugging Face listing itself use the Hy3 name across what appear to be different points in this scaling process, so exact version boundaries are not yet fully clear from public reporting.
Hy3 vs GLM-5.2 vs DeepSeek-V4-Pro: How the Numbers Compare
Putting Tencent’s open-source releases next to its two closest domestic rivals shows how differently the major Chinese labs are approaching model size versus efficiency. The table below uses only the parameter counts, licensing terms, and benchmark claims that have been publicly reported by named outlets for each model.
| Model | Total Parameters | Active Parameters | License | Reported Benchmark Note |
|---|---|---|---|---|
| Tencent Hy3 (Aug 28, 2026 preview) | ~770 billion | ~49 billion | Not yet detailed by Reuters | Targets software engineering, research, financial analysis |
| Tencent Hy3 (July 6, 2026 release) | 295 billion | 21 billion | Apache 2.0 | Comparable to GLM-5.2/DeepSeek-V4-Pro; beat GPT-5.5 on FrontierScience-Olympiad (per GIGAZINE) |
| Zhipu AI GLM-5.2 | 753 billion | Not fully disclosed | Open weight | Reference point cited by Tencent for Hy3 comparisons |
| DeepSeek-V4-Pro | 1.6 trillion | Not fully disclosed | Open weight | Reference point cited by Tencent for Hy3 comparisons |
| Tencent Hunyuan-Large (Oct 2024) | 389 billion | 52 billion | Open source | Largest open-source MoE model at time of release, per Tencent’s GitHub repo |
What stands out is that Tencent’s newest preview, at roughly 770 billion total parameters, now sits closer in scale to GLM-5.2 than its own July release did. That is a significant jump for a company that spent most of 2024 and early 2025 positioning its open models as smaller, more efficient alternatives to the largest Chinese systems. If the August preview holds up under independent benchmarking, it would suggest Tencent has decided the fight for developer mindshare requires playing at the same total-parameter scale as DeepSeek and Zhipu, not just matching their output quality with a leaner architecture.
CodeBuddy and WorkBuddy: Tencent’s Commercial Angle
The detail in Reuters’ report that has drawn the most attention from industry watchers is Tencent’s stated plan to integrate the new model alongside CodeBuddy and WorkBuddy, its coding assistant and workplace productivity products. That is a different commercial strategy from a lab that open-sources a model purely to build developer goodwill or recruit talent. By wiring the same open weights into paid, branded products, Tencent gets to have it both ways: developers who want to self-host or fine-tune the model can pull it from Hugging Face for free, while enterprise customers who want a managed, supported version can pay Tencent directly through CodeBuddy or WorkBuddy subscriptions.
This mirrors a pattern already visible in Tencent’s earlier Hy3 rollout. When the company launched the full July 6 version, it briefly offered free API access through OpenRouter for a two-week window before shifting to paid API pricing, according to a Tencent-affiliated social media post reviewed as part of this report. Tencent Cloud has since gone further on the WorkBuddy side specifically, announcing that Hy3 access through WorkBuddy will remain free to users globally through August 31, 2026, before any metered pricing kicks in on that product. That kind of promotional free period followed by metered API pricing is a common go-to-market tactic for open-weight model releases: it drives fast adoption and benchmark testing in the free window, then converts a portion of that traffic into paying API and subscription customers once it closes.
Licensing Terms and Where to Download Hy3
Tencent’s full July 2026 Hy3 release switched over to the Apache 2.0 license precisely on its July 6, 2026 launch date, according to reporting from DigitalApplied, GIGAZINE, and VentureBeat, with no field-of-use or geographic restrictions attached to the weights. That is a materially more permissive license than some competing Chinese releases have used, and it is one reason Hy3 has been picked up quickly by third-party inference providers and cloud platforms looking to offer it as a hosted option. Reuters’ report on the August 28 preview did not spell out the exact license terms for the new, larger version, so developers evaluating it for commercial use should check the license file attached to the specific Hugging Face repository before deploying it in a production environment.
The model weights for the new preview are available directly from the tencent/Hy3 repository on Hugging Face, alongside the Hy3-FP8 quantized version for teams with tighter memory budgets. Developers working with the earlier July release can also find the model mirrored on ModelScope, GitCode, and CNB, the pattern Tencent has followed with its other Hunyuan-family releases including Hunyuan-Large and the Hy3 GitHub repository.
pip install huggingface_hub
huggingface-cli download tencent/Hy3 --local-dir ./hy3-weights
huggingface-cli download tencent/Hy3-FP8 --local-dir ./hy3-fp8-weightsWhy Tencent Is Betting Big on Open-Source AI
Tencent’s open-source strategy is running against a very different backdrop than it was two years ago, when Hunyuan-Large first shipped. DeepSeek’s rapid rise through 2025 reset expectations across the entire Chinese AI industry about what an open-weight model could accomplish on a fraction of the training budget that OpenAI, Anthropic, and Google were reportedly spending. That shift pushed nearly every major Chinese tech company, Tencent, Alibaba, and Zhipu AI among them, to treat open releases as a core part of their AI strategy rather than a side project.
For Tencent specifically, open-sourcing Hy3 serves a dual purpose. It gives outside developers, researchers, and smaller Chinese companies a free, capable model to build on, which expands Tencent’s footprint across the broader Chinese AI ecosystem well beyond its own paid cloud customers. At the same time, seeding the model across Hugging Face, ModelScope, GitCode, and CNB simultaneously ensures Tencent’s technology gets evaluated and cited by international researchers and benchmark maintainers, which matters for recruiting AI talent and for how Tencent is perceived relative to DeepSeek, Alibaba’s Qwen team, and Moonshot AI’s Kimi line in the global open-source leaderboards.
Market Context: Tencent Holdings and the 0700.HK Story
Tencent trades on the Hong Kong Stock Exchange under ticker 0700.HK, and Reuters maintains ongoing market coverage of the stock alongside its news reporting on the company. As of publication, no specific share-price move tied directly to Friday’s Hy3 preview announcement had been reported by named financial outlets, so this article does not attribute any stock movement to the release. That said, AI product announcements have become a recurring catalyst for how investors read Tencent’s broader cloud and enterprise software business, since each new open-source release feeds directly into products like CodeBuddy and WorkBuddy that Tencent counts as part of its cloud and business services revenue.
The absence of a confirmed analyst reaction as of this writing is itself informative: it suggests markets are still digesting whether a preview-stage model release, as opposed to a full production launch, is a material enough event to move estimates for Tencent’s cloud or advertising businesses. Investors watching the stock will likely wait for either a full model release with published benchmarks or a formal product announcement tying Hy3 into a priced enterprise offering before adjusting expectations.
China’s Open-Source AI Strategy in a Restricted-Chip World
Tencent’s release lands inside a broader pattern that has defined Chinese AI development since US export controls began restricting access to the most advanced Nvidia data center GPUs. Rather than compete head-on with US labs purely on raw training compute, Chinese AI companies including Tencent, DeepSeek, Alibaba, and Zhipu AI have leaned into open-weight distribution as a way to maximize the reach and influence of models trained under real hardware constraints. Publishing model weights broadly, rather than gating access behind a single paid API, lets these companies build global developer mindshare even when their training clusters are smaller than what OpenAI, Anthropic, or Google can deploy.
None of the sources reviewed for this article included a direct statement from Tencent about US export controls or chip restrictions tied specifically to Friday’s release, so this article does not attribute any such statement to the company. The broader strategic context, however, is well documented across the industry: open weights spread faster and further than any single company’s compute budget could reach on its own, which is precisely why the strategy has become the default playbook across Chinese AI labs over the past 18 months.
Competitive Landscape: Tencent, Alibaba, DeepSeek, and Zhipu
The open-source coding and research model space in China is now genuinely crowded, and Tencent’s August 28 preview puts it back in direct competition with several well-established rivals rather than carving out a clearly separate niche.
| Company | Flagship Open Model | Primary Focus | Distribution Channels |
|---|---|---|---|
| Tencent | Hy3 (Hunyuan line) | Coding, research, financial analysis | Hugging Face, ModelScope, GitCode, CNB |
| DeepSeek | DeepSeek-V4-Pro | General reasoning, coding | Hugging Face, GitHub |
| Zhipu AI | GLM-5.2 | General reasoning, agentic tasks | Hugging Face, ModelScope |
| Alibaba | Qwen3.8-Max | General reasoning, multilingual | Hugging Face, ModelScope, GitHub |
What differentiates Tencent’s approach so far is the explicit tie-in to its own commercial software, CodeBuddy and WorkBuddy, rather than positioning the open model purely as a research contribution. DeepSeek’s releases have tended to emphasize raw benchmark performance and low API pricing over deep integration with a specific Tencent-style product suite, while Alibaba has leaned on Qwen’s broad multilingual coverage to court international developers building outside China. Tencent’s bet is that owning the full pipeline, from open weights to a branded coding assistant, gives it more durable revenue than benchmark bragging rights alone.
What This Means for Software Engineers and Enterprise Developers
For developers deciding whether to test Tencent’s new preview, the practical calculus mirrors what teams have already worked through with DeepSeek and GLM releases over the past year. A 770-billion-parameter MoE model with roughly 49 billion active parameters is not something most teams will run on a single workstation; it requires a multi-GPU inference setup or a hosted API, whether that is Tencent’s own infrastructure, OpenRouter, or a third-party cloud provider that picks up the weights once benchmarks confirm the model’s real-world performance.
Teams already running the July Hy3 release in production have an easier path: the FP8-quantized variant offers a way to test whether the scaled-up preview delivers a meaningful quality jump before committing more inference budget to it. Given that Reuters’ report did not include independently verified benchmark scores for the August 28 preview, engineering teams evaluating it now should treat any performance claims as provisional until third-party evaluators publish their own SWE-bench, HumanEval, or FrontierScience-Olympiad results against the new weights.
Predictions: Where Tencent’s Open-Source Push Goes From Here
Based on Tencent’s release pattern over the past two years and the trajectory of its Hunyuan and Hy3 lines, a few developments look likely in the coming months, though these remain analysis rather than confirmed fact.
- Tencent will likely publish formal benchmark numbers for the August 28 preview within weeks, following the same pattern it used with the July 6 Hy3 release, once the model moves from preview to full production status.
- CodeBuddy and WorkBuddy will probably see feature updates tied to the new model within the current quarter, given Reuters’ report that Tencent plans to integrate the release into those products directly.
- Rival Chinese labs, particularly DeepSeek and Zhipu AI, are likely to respond with their own scaled-up releases in the following months, continuing the pattern of rapid one-upmanship in total parameter counts that has defined the last year of Chinese open-source AI competition.
- Third-party inference providers such as OpenRouter and various cloud GPU marketplaces will likely add hosted access to the new Hy3 preview within days of stable weights becoming available, following the same fast-adoption pattern seen with the July release.
- Expect continued ambiguity in public reporting about exact model naming and version boundaries within the Hy3 line until Tencent publishes a formal technical report or model card clarifying how the August preview relates to the July release.
Historical Context: How Chinese Open-Source AI Got Here
It is worth stepping back to see how quickly this landscape has moved. Tencent’s Hunyuan-Large, at 389 billion total parameters, was described as the largest open-source MoE model in the industry when it launched in October 2024. Less than two years later, that figure looks almost modest next to DeepSeek-V4-Pro’s reported 1.6 trillion parameters and GLM-5.2’s roughly 753 billion. Tencent’s own trajectory, from Hunyuan-Large’s 389 billion, to the April 2026 Hy3 preview, to the July 295-billion-parameter full release, to Friday’s reported 770-billion-parameter preview, tracks that same industry-wide scaling curve almost exactly.
That pace of iteration is itself a story: fewer than eight months separate the April Hy3 preview from Friday’s much larger follow-up preview, a release cadence that would have been unusual for a frontier model lab even two years ago. It reflects both the intensity of competition among Chinese AI labs and the fact that mixture-of-experts architectures let companies scale total parameter counts substantially without a proportional increase in inference cost, since only a fraction of the network activates per query regardless of how large the full model gets. As one independent technical breakdown of the July Hy3 release put it, the architecture choice is what lets Tencent compete on capability without matching rivals parameter-for-parameter.
Frequently Asked Questions
What is Tencent Hy3?
Hy3 is the name Tencent has used across a series of open-weight AI model releases in 2026, built on a mixture-of-experts architecture and aimed at coding, research, and financial analysis tasks. The August 28, 2026 release is a preview version that Reuters reported carries roughly 770 billion total parameters with about 49 billion active per request.
Is Tencent Hy3 free to use?
The model weights are published openly on Hugging Face, which means developers can download and self-host the model without a licensing fee, subject to the terms of whichever license Tencent attaches to the specific release. Tencent’s July 6, 2026 full Hy3 release shipped under the Apache 2.0 license; the exact license for the August 28 preview was not detailed in Reuters’ initial report.
How does Hy3 compare to DeepSeek-V4-Pro and GLM-5.2?
Tencent’s July 2026 Hy3 release, at 295 billion total parameters, was described by the company as delivering scores comparable to GLM-5.2 (roughly 753 billion parameters) and DeepSeek-V4-Pro (roughly 1.6 trillion parameters), according to a report from GIGAZINE. The newer August preview, at a reported 770 billion parameters, moves Tencent’s flagship open model much closer in scale to GLM-5.2.
What is CodeBuddy?
CodeBuddy is one of Tencent’s coding assistant products. Reuters reported that Tencent plans to integrate the new open-source model alongside CodeBuddy and WorkBuddy, tying the open release directly into paid products the company already offers to developers and enterprise customers.
Where can I download the Tencent Hy3 model?
The model is available on Hugging Face under the repository tencent/Hy3, with a quantized Hy3-FP8 variant also published. Earlier Hy3-family releases have also been distributed through ModelScope, GitCode, and CNB.
Did Tencent’s stock move after the Hy3 announcement?
No specific share-price reaction tied to the August 28, 2026 announcement had been reported by named financial outlets as of publication. Reuters maintains ongoing market coverage of Tencent under ticker 0700.HK, but this article does not attribute any stock movement to the release since none was confirmed in available reporting.
Does Hy3 target only coding tasks?
No. Reuters specifically described the August 28 preview as aimed at software engineering, research, and financial analysis, indicating Tencent is positioning the model for broader analytical work rather than code generation alone.
Is this the same model as the April 2026 Hy3 preview?
Not exactly. The April 2026 preview that Bloomberg reported on, the July 6, 2026 full release (295 billion parameters), and the August 28, 2026 preview (reported at roughly 770 billion parameters) all use the Hy3 name but appear to represent different stages of scaling. Tencent had not published a technical report clarifying the exact relationship between these versions as of this article’s publication.
Related Coverage
- DeepSeek V4 vs R1 vs V3.2: Peak Prices Surge 355% [2026]
- Kimi K3 vs Qwen3.8-Max vs GLM-5.2: $10.60 Gap [2026]
- Nvidia Reportedly Buys Hugging Face for $12.9B [2026]
- Best Open Source LLM [2026]: DeepSeek V4 Hits 80.6%
- Meta Muse Glimmer: 30B AI Model Fits in 20GB VRAM [2026]
- Local AI Beats ChatGPT on Reliability, 20x Cheaper [2026]
- Best AI Models 2026: Full Comparison Hub
