Tencent Hy4 launches Nov 4: weights due 'in the coming days', API price and tests
Tencent's full Hy4 adds vision and computer use and is on sale by API now, but only the August preview weights are on Hugging Face so far.
· 7 min read
Written by AI from the sources linked below. How we use AI
On this page
Tencent announced Hy4 on 4 November 2026, the full release of the large language model it first showed as "Hy4 Preview" in August. Tencent's announcement adds visual understanding and computer use, and says model weights and system cards will follow "in the coming days". At the time of writing the only Hy4 weights on Hugging Face are the August preview. So developers can call the new model through paid services today, but cannot yet download it.
Key points
- Tencent announced Hy4 on 4 November 2026 and says it plans to publish the weights and system cards "in the coming days"; Hy4 Flash is due "soon".
- Tencent lists API pricing of about 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.
- In a Tencent-run blind evaluation of 579 financial tasks judged by 213 professionals, Hy4 scored 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3.
- The only Hy4 weights public today are the August Hy4 Preview (770B parameters, 49B active per token, 1M-token context, Apache 2.0).
- All performance figures in this article come from Tencent and have not been independently reproduced.
What happened
Tencent's own announcement, published on 4 November, presents Hy4 as joining "the ranks of the world's leading open-source AI models". The same text says Tencent Hy "plans to release model weights and system cards in the coming days" and that a smaller Hy4 Flash is coming "soon". That wording matters: on the day of the announcement, a search of Tencent's Hugging Face organisation through the public API returned two Hy4 repositories, tencent/Hy4-preview and tencent/Hy4-preview-FP8, both created on 27 August 2026. No repository for the final model was listed. Until the weights appear, "open-source" describes Tencent's intention and its earlier preview, not the model announced this week.
What is new, according to Tencent, is a set of capabilities the preview lacked. Hy4 adds visual understanding and computer-use abilities for the first time, and Tencent says it was trained on professional software such as Blender and CAD. The company also says it trained on defensive security data and excluded offensive exploitation material, and it describes the model's biosafety capabilities in open-ended experimental design as more limited than those of frontier models.
The model is available through Tencent Cloud TokenHub and Tencent's own products (WorkBuddy, CodeBuddy, Yuanbao, Dayuan and ima). Tencent says the API is also reachable through OpenRouter, Hermes and OpenCode. Two promotions run alongside the launch: WorkBuddy and CodeBuddy are free from 4 to 15 November, and Hy4 is 50% off standard pricing on TokenHub through 31 December 2026.
Background
Tencent's Hy line has moved quickly this year. In late August (the repositories were created on 27 August) the company put the Hy4 Preview weights on Hugging Face under the Apache 2.0 licence. The model card lists a mixture-of-experts design with 770 billion parameters in total and 49 billion active for each token, 78 layers (the first dense, the other 77 mixture-of-experts), 256 routed experts plus one shared expert with the top eight activated per token, and a context length of one million tokens. A built-in multi-token prediction layer adds 10 billion parameters (0.7 billion active) and is used for speculative decoding.
On 29 October, at the Tencent Global Digital Ecosystem Conference, senior executive vice president Dowson Tong said Tencent planned to release Hy4 and Hy4 Flash soon with visual understanding, as TechNode reported, citing Chinese outlet IT Home. TechNode's account says the full model is meant to do better on tasks that need sustained reasoning and execution, such as real-world productivity workflows, and that Hy4 Flash would focus on cost efficiency while keeping visual understanding and agent abilities. The 4 November announcement is the delivery of the first half of that plan.
For readers tracking open-weight models, the context is a crowded field. We have covered Kimi K3.1 Pro's open weights and licence terms and Revolut's report that self-hosted GLM cut its AI coding costs. Tencent chose those same two rivals, Kimi K3.1 Pro and GLM-5.3, as comparison points in its own evaluation.
The numbers
Tencent published several results for the new model. The headline one is a blind evaluation of financial tasks: 213 professionals assessed 579 tasks, and Hy4 averaged 3.04.
The gap to DeepSeek-V4.1-Flash-1017 is 0.07 points, small enough that without published variance or rater details it should not be read as a clear win. The gap to Kimi K3.1 Pro and GLM-5.3 is larger, 0.36.
Other figures from the announcement, all from Tencent:
- WorkBuddy, Tencent's agent product: an average score of 83.18 and a 96.15% pass rate, ranking first in nine of twelve scenarios in mainland China.
- Compared with Hy4 Preview, the hallucination rate fell from 12.95% to 7.2% in Tencent's internal evaluations.
- Also against the preview: 19.1% better on the GameDev benchmark and 9.8% better on GameLogic, and a 12.03 percentage-point improvement on iRaB (from Rabyte).
For context, the preview's model card lists results such as 92.3 on GPQA Diamond, 85.4 on Terminal-Bench 2.1 and 65.7 on SWE-Bench Pro. Those are preview numbers; Tencent's announcement does not restate them for the final model in the text we could read, so we do not assume they carry over.
Pricing is unchanged from the preview: about US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens. OpenRouter listed the preview at a similar level (about US$0.83 and US$2.50 per million tokens), according to eesel.ai's preview write-up.
How it works
The final architecture has not been published, because the system card has not been released. What we can say comes from the preview. A mixture-of-experts model stores many specialist sub-networks ("experts") but sends each token through only a few of them, which is why a 770-billion-parameter model computes with roughly 49 billion parameters per token. That keeps per-token compute closer to a mid-sized dense model, but the full set of weights still has to sit in memory. The preview's deployment instructions show how heavy that is: the vLLM and SGLang examples on the model card use eight GPUs (tensor parallel size 8), and eesel.ai estimates about 1,848 GB for the BF16 weights and about 924 GB for the quantised build. The model card itself does not state a memory requirement, so treat those two figures as a third-party estimate.
If the final weights follow the preview's pattern, the likely route will be the same: serve the FP8 build with vLLM or SGLang across eight accelerators, with speculative decoding enabled. That is an inference from the preview, not something Tencent has confirmed. The preview card recommends temperature 0.9 and top_p 1.0, and lets you pass reasoning_effort: "no_think" for direct answers. Check the final card for changes.
What it means
For developers who want to try Hy4 now, the hosted API is the only route. Costs depend on how many tokens the model spends reasoning before it answers, so test with your own workload. The free windows on WorkBuddy and CodeBuddy end on 15 November.
For teams that need self-hosting, nothing changes yet. The preview weights are usable today under Apache 2.0. The final licence has not been stated in the announcement, so do not assume it matches. Plan for a hardware footprint in the hundreds of gigabytes at least.
For the wider market, Tencent is following the preview with a hosted launch first and weights later. The risk for readers is confusing the stages: "preview weights are open", "final model is available by API" and "final weights are downloadable" are three separate facts, and only the first two are true today.
What we don't know yet
- When the final weights will actually be published, and under which licence. Tencent says "the coming days" and gives no date.
- The final model's parameter count, architecture changes and context length. Only the preview's are public.
- Whether Hy4 Flash's weights will be released alongside it.
- How the model performs on independent, public benchmarks. Every number above is Tencent's. The evaluation is described by the company as its own, and rater details beyond the headcount are not published in the announcement.
- Whether the visual and computer-use abilities hold up outside Tencent's own products.
What to watch next
Watch Tencent's Hugging Face organisation for a final Hy4 repository and its model card, and for the system card, which should state the licence and safety findings. Watch for Hy4 Flash, for third-party results from independent evaluators, and for the end of the 50% TokenHub discount on 31 December 2026.
Sources
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
Ripple-65B Preview: 65B open-weight coding model, 80.5 on TerminalBench 2.1, Apache 2.0
Ricursive Intelligence says its dense 65B model beats US open models up to 15 times larger on agentic coding for $32 million; the benchmarks are self-reported.
audiotale.net team7 min read
Revolut says self-hosted GLM cut AI coding costs 90% for 2,000 engineers on 56 B200s
A Revolut infrastructure lead told AI Engineer that self-hosting open-weight models cut coding inference costs by about 90 percent; the reports disagree on the model.
audiotale.net team6 min read
Kimi K3.1 Pro open weights: 2.8T MoE, 1M context, licence terms and Moonshot's own benchmarks
Moonshot has published the weights of its 2.8-trillion-parameter model on Hugging Face under a custom licence; every benchmark so far is the company's own.
audiotale.net team6 min read