Hy4 Preview: Tencent Hunyuan Release, Pricing and Access

What is Hy4 from Tencent Hunyuan?

Hy4 is the next model in Tencent Hunyuan's Hy family, teased on Nov 3, 2026. LMArena, the leaderboard operator, posted that "Hy4 from @TencentHunyuan is coming soon." Tencent has not published a full announcement yet.

One framing note for AdGraam readers: everything measured so far is agent work in code, work and chat. Nothing published points to video or image generation, and Tencent has not said what the final Hy4 will cover. For ad teams, the preview matters mainly as a cheap, long-context model for the work around creative: scripts, briefs, research, tagging and automation.

A preview version already exists. The LMArena post references last month's "Hy4 preview launch," and a GitHub repository named Hy4-preview exists under the Tencent-Hunyuan organization.

Hy4 release date and availability

Tencent has not announced a Hy4 release date. The only timing signal is the LMArena post on Nov 3, 2026, which says the final release is "coming soon" and promises to show where it lands in the arena.

Tencent Hunyuan's own account also posted a teaser titled "4 Real Productivity," quoted in a post from @laopai1990. That post says a Flash version may ship at the same time and that the teaser's visual style resembles a Kimi K3.1 Pro promo. Both points are the poster's opinion. Tencent has not confirmed a Flash model.

The dates for the preview do not line up cleanly. LMArena calls it a launch from "last month," while the preview's GitHub repository shows commits dated Aug 28, 2026. The repository, as mirrored on a third-party page, shows five commits and a fine-tuning folder. It is not clear from what has been published which date counted as the public launch.

Hy4 preview benchmarks: what Agent Arena shows

In Agent Arena, the Hy4 preview ranks #2 among open models and #10 overall, according to LMArena. The figures come from 14.5K+ real-world agent sessions and show a +5.4% net improvement, behind only Kimi K3 (Max) at +6.6%.

LMArena also reports cost: roughly $0.26 per median task for Hy4 preview versus $0.80 for Kimi K3 (Max), about 68% lower. It says the preview sits on the Agent Arena Pareto frontier.

By signal, LMArena lists:

• Confirmed Success: 12%, explicit user feedback that the task worked
• Praise vs. Complaint: 9%, implicit sentiment in user reactions
• Bash Recovery: 7.5%, recovery from CLI errors
• Tool Hallucination: no issues reported

Among open models, it is #2 in Code and Work and #3 in Chat. These are arena results from LMArena, a leaderboard operator rather than Tencent, and they measure agent tasks, not marketing output quality.

Hy4 preview pricing and context window

A pricing roundup published Oct 30, 2026 lists Hy4-preview at $0.834 per 1M input tokens and $0.042 per 1M cached input tokens, with a 1,000K token context window. The article takes its rates from AM Compute and cites Tencent Cloud documentation for the window. Output pricing is listed as not verified.

A long paper scroll beside a brass scale, representing a very large context window and token pricing

The preview costs more per input token than Hy3 but offers roughly four times the context. The same roundup says rates can differ by region and billing console, so check your own console before budgeting. Final Hy4 pricing is not announced.

On license, the Hy4-preview repository shows a badge reading Apache 2.0. Terms for the final model have not been published.

How to use Hy4 in an ad creative workflow

If Hy4 ships as previewed, the likely fit is the text side of production: long-context analysis and agent tasks rather than pixels. That is an inference from the arena categories (code, work, chat) and the 1,000K window, not a Tencent claim.

A desk with ad mockups and an abstract workflow diagram representing ad production planning

Practical uses to test once it is available:

• Feeding a full product catalog, review dump or competitor ad transcripts into one prompt to extract angles and hooks.
• Drafting and localizing scripts at volume, where lower per-task cost matters.
• Running agent steps that tag, rename and organize large creative libraries.

Pair any text model with image and video generators for the actual assets. In AdGraam, the Playground lets you run and compare models side by side, and Flow Builder chains steps into repeatable ad workflows. For the hook side of testing, see our hook rate and hold rate benchmarks.

On the visual side, Tencent's own image model is separate. Hy Image 3.5 Preview was released Sep 22, 2026, according to a gateway listing, and supports any width and height from 256 to 8192 pixels up to a 4096x4096 area. It generates from text and edits from one or more reference images.

Hy4 vs other AI models

Against Kimi K3 (Max), the only open model ahead of it in Agent Arena, Hy4 preview trails by about one percentage point (+5.4% versus +6.6%) at roughly a third of the median task cost, per LMArena. That is the only head-to-head comparison available.

No comparison against closed models beyond the #10 overall rank has been published. Treat the cost figure as an arena median, not a price list.

What to watch next

Watch for three things: a Tencent announcement with a date, the LMArena placement of the final Hy4, and whether a Flash tier appears with its own pricing.

Also unresolved are output token rates, the final license, and whether Tencent adds multimodal input. Until then, treat the preview numbers as the best available guide. Since Hy4 may ship soon, put together a small set of your own briefs and scripts now, so you can run it against your current model the day it launches.