Tencent Launches Hy4 With Vision and Computer Use at Half Price. What About Arabic?

On 4 November 2026, Tencent released Hy4, the first model in its Hy line that can both read images and operate a computer. API pricing matches the preview release at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens. You can use Hy4 through Tencent Cloud and OpenRouter, alongside integrations in CodeBuddy, WorkBuddy, Yuanbao, and ima.

What changed, and who it is built for

Hy4 has 770 billion parameters, according to TechNode. Compared to the Hy4 preview from 27 August 2026, the update expands pre-training and post-training, adding native vision for fine details and complex charts. It introduces computer use across command-line and graphical interfaces, with targeted training on Blender, CAD, and multi-application tasks. On OpenRouter it takes text and images, holds a context of about one million tokens (1,048,576) and reasons by default. Tencent also claims gains in instruction following and deep research.

Tencent aims it at four kinds of work: finance (developed with Tencent Financial Technology), office work via WorkBuddy and WeCom (chats, emails, docs, calendars, notes), frontend and game development, and scientific research.

What it costs

The list price is $0.834 input and $2.501 output per million tokens, but almost nobody pays it right now. OpenRouter currently shows a 50% discount, $0.417 input and $1.25 output, without an end date on the listing. Vercel AI Gateway says its 50% discount runs through 31 December 2026. TechNode reported a matching 50% discount on Tencent Cloud TokenHub through 31 December, alongside free quotas in WorkBuddy and CodeBuddy from 4 November to 15 November. A cheaper, faster Hy4 flash model is planned “in the near future,” without a confirmed date.

Where it falls short

In its research announcement, Tencent claims Hy4 ranks “among the strongest open-source models for the first time,” but admits “a notable gap remains” against leading proprietary models in research maths, long-horizon coding, and advanced visual Q&A. Claude Opus 5.5 leads Hy4 across all twelve benchmarks Tencent displayed.

Independent numbers are less flattering. As reported by The Decoder, Hy4 scored 35 on the Artificial Analysis Intelligence Index (median 30), placing second among Chinese models behind Kimi K3.1 Pro (46) level with Grok 4.7 and roughly with Anthropic’s Haiku 5.5. Worse, on the AA-Omniscience test, Hy4 fabricates an answer 89% of the time when it lacks information instead of admitting it.

Open weights: promised, not dated

Tencent says it “plans to open-source” the weights, with a system card and vLLM and SGLang support, but gives no date and no licence. Its predecessor Hy3 (295B parameters) shipped under Apache 2.0 in July. Since then, The Decoder notes, China has drafted export-control rules in October that could cover top model weights. Tencent also says it filtered offensive hacking material out of training and that the model scored below its own biosecurity risk threshold. Until the weights appear, Hy4 is a cloud model like any other.

Does it handle Arabic? Test it yourself

Tencent publishes no language list, and Arabic is never mentioned by Tencent, OpenRouter, Vercel, or tech reporting. English benchmarks will not answer that, so spend 15 minutes on OpenRouter. It is the one platform outside Tencent’s own products that the announcement names, and you can pay there without opening a Tencent account.

Compare Hy4 side by side against DeepSeek or Kimi K3.1 Pro across three tasks:

• Document vision: Upload a photo of an Arabic invoice or financial chart to test OCR and numerical accuracy.
• Dialect comprehension: Prompt the model to summarise a conversational Gulf-dialect message into standard prose.
• Tabular reasoning: Submit an Arabic spreadsheet table with an analytical query to evaluate right-to-left data handling.

Our take: who should try it now

Teams building multimodal agents, GUI automation, or visual document pipelines should try Hy4 now on OpenRouter while the launch discount lasts; at half price it is a cheap second opinion next to DeepSeek or Kimi. Anyone who needs reliable Arabic answers in production, factual work without web search, or weights to host privately should wait: the 89% figure alone is reason to keep it away from unchecked facts, and the weights have no date.