Tencent announced Hy4 on November 4, 2026 with vision, coding and computer use. Its benchmark results are Tencent’s own, and the full weights had not been published by November 12.
Tencent announces Hy4
Tencent introduced Tencent Hy4 on November 4, 2026. Tencent’s announcement says the model brings visual understanding “for the first time”, with gains in instruction following, long-context comprehension and long-horizon agentic tasks. Tencent also says Hy4 can use command-line and graphical interfaces, another capability it describes as new.
Against its earlier Hy4 Preview, Tencent reports a hallucination rate that fell from 12.95% to 7.2% in internal evaluations. It also reports a 12.03 percentage-point improvement in the internal Investment Research Agent Benchmark (iRaB), which it says was conducted by the financial AI company Rabyte (Tencent). Both figures are Tencent’s own.
Tencent titles its announcement as joining the ranks of the world’s leading open-source AI models, and its text calls Hy4 the top-performing open-source model in its own financial evaluation. That framing arrived before the model’s weights did. Tencent says it plans to release weights and system cards “in the coming days”. By November 12 the full Hy4 weights had not appeared, a point the section on running Hy4 locally covers. The term open-weight applies today only to Hy4 Preview, which is a different model.
Tencent calls the family Hy and its team the Tencent Hy team. Its repositories still carry the older name, for example the Tencent-Hunyuan GitHub organization and the hunyuan tag on the Hy4 Preview model page.
What Tencent says the model can do
| Capability | What Tencent says | Example Tencent gives |
|---|---|---|
| Vision | Fine-grained image processing, complex chart analysis and mathematical visual reasoning, new in this model | Visual understanding combined with tool calling |
| Computer use | Works with command-line and graphical interfaces, also new | Specialised training on professional software such as Blender and CAD |
| Coding and games | Complex business logic, state management and front-end and back-end orchestration | Integration with Unreal Engine, Unity and Godot, with 19.1% and 9.8% gains over Hy4 Preview on the GameDev and GameLogic benchmarks |
| Instruction following, long-context comprehension and long-horizon agent tasks | Listed by Tencent as three separate improvements | No separate figure is given for these three |
| Workplace | Better results in Tencent’s WorkBuddy product | Average score 83.18, pass rate 96.15%, first in nine of twelve scenarios in the mainland Chinese market, all measured by Tencent. Through WeCom via WorkBuddy, Hy4 tracks project history across chats, emails, documents, calendars and meetings |
| Financial analysis | Trained on workflows for fund research, securities analysis and quantitative investment strategies | Thousands of specialised workspaces and tens of thousands of tasks used in training |
Source: Tencent announcement. These are capabilities Tencent announces, not performance measured on independently selected workloads. Tencent also states that in its internal expert reviews Hy4’s capability in open-ended biological experimental design is more limited than that of frontier models.
Vendor-reported evaluation results
Tencent says the Tencent Hy team ran a blind financial-task evaluation in partnership with Tencent Financial Technology. According to Tencent, it involved 213 financial professionals, of whom 153 were experts from leading external institutions, and about 70% had more than five years of industry experience. The evaluation covered 579 real-world financial tasks.
| Model | Overall score reported by Tencent |
|---|---|
| Hy4 | 3.04 |
| DeepSeek-V4.1-Flash-1017 | 2.97 |
| Kimi K3.1 Pro | 2.68 |
| GLM-5.3 | 2.68 |
Source: Tencent announcement. Tencent does not state the scale of these scores, and the announcement does not publish the task set or the scoring rubric. The model’s developer ran the test, in partnership with Tencent Financial Technology. One comparator, DeepSeek-V4.1-Flash-1017, carries Flash in its name, which suggests a smaller tier than the flagship models beside it. The scores are therefore not an independent ranking of model intelligence. The announcement also describes other internal and partner evaluations of workplace, scientific and safety tasks. TechiesJournal’s guide to how to read AI model claims covers what to ask of any such number.
What independent testing shows
Vals AI, a company that runs its own model benchmarks, is the one independent measurement of Hy4 found when sources were read. Its Hy4 page reports 59.35% on the Vals Index, plus or minus 1.08, which places Hy4 15th of 36 models. The Vals Index leaderboard, updated November 9, 2026, describes the index as agentic performance across finance, coding, legal and tax tasks.
| Model | Vals Index score | Rank on Vals leaderboard |
|---|---|---|
| Kimi K3.1 Pro | 62.52% | 10 |
| DeepSeek V4.1 Flash 1017 | 61.26% | 11 |
| GLM 5.3 | 59.85% | 14 |
| Hy4 | 59.35% | 15 |
| Hy4 Preview | 53.21% | 26 |
Source: Vals Index leaderboard, read on November 12, 2026. The Hy4 page also reports a cost of $1.720 per test on the Vals Index and a Vals Finance Index of 58.05%, 15th of 39 models.
Three of the models Tencent compared against sit above Hy4 on this index. By our arithmetic, the gap is 0.50 points to GLM 5.3, 1.91 to DeepSeek V4.1 Flash 1017 and 3.17 to Kimi K3.1 Pro. The gap to GLM 5.3 is smaller than the margin Vals prints for Hy4. Four caveats apply:
- Vals tests different tasks from Tencent’s blind finance evaluation, and no independent test of Tencent’s financial claim was found.
- Vals is a single benchmark company, so its ranking is one more data point and not a final verdict.
- Vals’s own pages are not fully consistent. A note on the Hy4 page dated November 4 gives GLM 5.3 as 59.73% and Hy4 as 13th of 33, while the leaderboard updated November 9 gives 59.85% and 15th place.
- Vals labels Hy4 as an open-weight model, a label we could not verify because the full weights had not been published.
For the practical side of the other three models, see TechiesJournal’s earlier article on whether teams can use DeepSeek, GLM and Kimi open-weight models.
Availability and announced pricing
Tencent says Hy4 is available across Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom’s AI Assistant (Dayuan), ima and more, and that its API is being made available through third-party platforms such as OpenRouter, Hermes and OpenCode. Integration efforts for QQ, Honor of Kings, Game of Peace, WeRead and Tencent News are under way, which means those are not available yet (Tencent). Of the third-party platforms, only OpenRouter was checked: it listed tencent/hy4 on November 12, 2026. Hermes and OpenCode remain Tencent’s statement.
Tencent gives approximate prices of US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens, the same as Hy4 Preview. Tencent says TokenHub offers Hy4 at 50% off its standard pricing through December 31, 2026, and that WorkBuddy and CodeBuddy are free from November 4 to 15 (Tencent). Tencent does not say in words whether the US$0.90 figure is before or after the discount, so this article does not apply the discount to it.
| API usage, per million tokens | Tencent’s approximate USD figure | OpenRouter listing, Tencent endpoint, observed November 12, 2026 |
|---|---|---|
| Input | $0.90 | $0.417 |
| Output | $2.70 | $1.2505 |
| Cached input | $0.045 | $0.021 |
Sources: Tencent announcement and OpenRouter’s Hy4 page. OpenRouter’s figures are its own listing and can change. Its endpoint data for the Tencent provider also carries a discount value of 0.5, and its data names the cached figure input_cache_read. These rates describe API usage. They do not establish the operating cost or hardware needs of a self-hosted deployment.
Model names and identifiers
| Name | Where it appears | Status on November 12, 2026 |
|---|---|---|
| Hy4 | Tencent announcement | Announced November 4, 2026 and offered through Tencent products and hosted APIs |
tencent/hy4 | OpenRouter | Listed with text and image input, a 1,048,576-token context window and 65,536 maximum output tokens |
tencent/Hy4-preview and tencent/Hy4-preview-FP8 | Hugging Face | Published August 27, 2026 under Apache-2.0 |
| Hy4 Flash | Tencent announcement | Tencent’s heading says “Recently Released” while its text says “scheduled for release soon”. It was not found in OpenRouter’s model list |
When a team writes down a model identifier, it should record which row it means. The hosted Hy4 and the downloadable Hy4 Preview are different models.
Can Hy4 be run locally?
Three dated facts answer the question as of November 12, 2026.
The full Hy4 weights were not public. Tencent said on November 4 that weights and system cards would follow in the coming days. Eight days later, the tencent organization on Hugging Face listed Hy4 Preview and Hy4 Preview FP8 but no full Hy4 repository. The Hugging Face address that OpenRouter gives for Hy4, tencent/Hy4, returned an authorisation error without sign-in. The Tencent-Hunyuan/Hy4 repository on GitHub returned not found, as did the equivalent address on ModelScope’s API. We found no Hy4 system card, and Tencent’s announcement states no licence for Hy4.
Hy4 Preview and its FP8 build are published under Apache-2.0. The Hy4 Preview model card describes a mixture-of-experts model with 770B total and 49B activated parameters and a 1M-token context length. It includes deployment sections for vLLM and SGLang with recipes, a finetuning section and a FP8 build. Its FP8 vLLM example sets tensor parallelism to 8. The same card lists known limitations, among them spending longer than necessary reasoning through complex tasks. The code is also on GitHub.
Hardware needs for the full Hy4 are not documented. Tencent describes Hy4 as a significant improvement over Hy4 Preview, and Vals AI scores them differently (59.35% against 53.21% on its index). The Preview’s serving recipes are therefore a starting point for planning, not documentation for the November 4 model.
Teams that need self-hosting should check Tencent’s Hugging Face organization for the full Hy4 repository, its licence file and its system card before committing to any capacity plan.
What developers should know
Hy4 can be used today through Tencent’s hosted products and OpenRouter at the prices above, and only Hy4 Preview can be downloaded and served. Tencent’s headline results are its own, and the one independent index found places Hy4 close to, but below, the three models Tencent compared it with. Test it on your own tasks, and recheck the weights, licence, system card, prices and the TokenHub discount before relying on any of them.
References
- Tencent: Hy4 launch announcement, November 4, 2026
- Tencent models on Hugging Face
- Hy4 Preview model card and Apache-2.0 licence
- Hy4 Preview FP8 model
- Tencent-Hunyuan Hy4 Preview repository on GitHub
- OpenRouter: Tencent Hy4
- Vals AI: Hy4 model page
- Vals AI: Vals Index leaderboard
Sources were read on November 12, 2026. Scores, prices and availability are Tencent’s, OpenRouter’s and Vals AI’s, and can change.