# Tencent Hy4 Explained

> Tencent's flagship model, released in full on November 4, 2026: vision and computer use added to the text-only Hy4 preview from August 2026. API cost calculator at list and 50 percent off prices, vendor-reported benchmark table with conditions, access checker, and setup steps for the OpenRouter API. The full weights are promised but not published; the API and Tencent apps are the only ways to use the full model as of November 5, 2026. Independent explainer, not affiliated with Tencent. Updated 2026-11-05. Full page: https://lab.awaara.cn/tencent-hy4/

## What it does

- Cost calculator: enter input tokens per run, cache hit percent, output tokens per run, and runs per day. It splits input into cached and uncached parts, applies the selected price set, and shows cost per run, per day, and per 30 days, with a line by line breakdown. Presets: 1M token batch (800,000 input, 200,000 output, no cache, $1.17 at list price, $0.58 discounted), agent day (300,000 input, 70 percent cache, 15,000 output, 20 runs), short chat reply.
- Access checker: pick what you need (Tencent app, API, weights, local run, Hy4 Flash, EU in-region routing) and get the status Tencent published on November 4, 2026. Weights, local runs, Hy4 Flash, and EU in-region routing are not available; Tencent apps and the API are.
- Benchmark explorer: 7 rows from Tencent's 34 row appendix, filterable by computer use, coding and agents, finance, and internal tests. All scores are Tencent's own measurements; 20 of the 34 benchmarks are Tencent internal tests.
- Setup steps: OpenRouter endpoint, model id tencent/hy4, and the 1,048,576 token context and 65,536 token completion limits from the OpenRouter listing. Everything on the page runs locally in the browser. Nothing is sent to a server.

## Facts

- Release: November 4, 2026. Preview launched August 28, 2026, text only, with weights under Apache 2.0. Hy3, the previous generation, launched in April 2026.
- What is new: vision (fine-grained image detail, complex chart analysis, visual math) and computer use (command-line and graphical interfaces, training on Blender and CAD tools and on long tasks across several apps). Pricing is unchanged from the preview.
- Scale, with the source kept apart: the full release announcement does not restate the parameter count, active parameters, or context length. Preview figures: 770 billion total parameters, 49 billion active per token, mixture of experts, about 1 million token context. OpenRouter listing for the full model: 1,048,576 token context, 65,536 token completion limit. Cite preview figures as preview figures.
- Pricing, list: $0.834 per million uncached input tokens, $0.042 per million cached input tokens, $2.501 per million output tokens. Yuan list price: 6, 0.3, and 18 per million tokens. Tencent's English press release rounds the list price to about $0.90, $0.045, and $2.70.
- Pricing, discounted: 50 percent off, $0.417, $0.021, and $1.251. Tencent Cloud TokenHub discount runs through December 31, 2026. OpenRouter shows the same discount with no end date stated in coverage checked. WorkBuddy and CodeBuddy are free from November 4 to 15, 2026, with a limited free usage quota.
- Benchmarks (all Tencent-reported, from the 34 row appendix): OSWorld 2.1: Hy4 78.9, Claude Opus 5.5 81.8, Kimi K3.1 Pro 79.6, GPT-6.1 Sol 75.2. Terminal-Bench 4.0: 44.4, 66.4, 66.8, 55.5. ProgramBench: 30.0 vs Claude Opus 5.5 at 56.5. APEX-Agents v1.1: 75.2 vs 74.8. Finance blind evaluation (579 tasks, 213 professionals): Hy4 3.04, DeepSeek-V4.1-Flash 2.97, Kimi K3.1 Pro 2.68, GLM-5.3 2.68. Internal tests: Hy-FrontierMath T5 13.0 vs GPT-6.1 Sol 97.5 and Claude Opus 5.5 91.3; ProofBench 5.5 vs 70.5 and 42.2. Gains over the preview: GameDev up 19.1 percent, GameLogic up 9.8 percent, iRaB up 12.03 percentage points, WorkBuddy average 83.18 with a 96.15 percent pass rate and first place in 9 of 12 scenarios.
- Hallucination figures, two pairs reported: the English press release says 12.95 percent in the preview to 7.2 percent, with unsupported-assumption errors down from 23.1 percent to 5.6 percent. Coverage of the Chinese release post reports 15.76 percent to 5.45 percent. Do not merge the pairs; they come from different posts and evaluations.
- Safety (Tencent-reported): cybersecurity training limited to defensive material; exploitation, reverse engineering, and penetration testing material removed; Xuanwu Lab and Keen Security Lab audits found no pronounced offensive risk in the areas checked. Hy-BioExperimentBench: 30.4 out of 100, below Tencent's high-risk threshold. The system card with fuller safety results is promised with the weights.
- Weights and license: full-model weights not published as of November 5, 2026. Tencent says it plans to release them and a system card in the coming days, with vLLM and SGLang support. No date and no license stated. Preview weights (tencent/Hy4-preview and FP8 build) are on Hugging Face under Apache 2.0.
- Hy4 Flash: planned faster and cheaper version, described as delivering greater cost efficiency. No date, specification, or price published as of November 5, 2026.
- Science models open-sourced the same day: HyCell (single-cell representations) and HyPDE (physics fields), built with Hy4 and the Hyra 2.0 harness. Separate models, not Hy4 itself.
- Access: Tencent Cloud TokenHub, OpenRouter (model id tencent/hy4, Tencent Cloud the only listed provider, no EU in-region routing), Hermes and OpenCode, and Tencent apps Yuanbao, WorkBuddy, CodeBuddy, ima, and WeCom assistant Dayuan.

## Sources

- Tencent, Tencent Launches Hy4: https://www.tencent.com/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models/
- OpenRouter, Hy4: https://openrouter.ai/tencent/hy4
- Unite.AI, Tencent Releases Full Hy4 Model With Vision and Computer Use: https://www.unite.ai/tencent-releases-full-hy4-model-with-vision-and-computer-use/
- MadRobot, Tencent Hy4 is out: https://madrobot.blog/2026/11/04/tencent-hy4-release-computer-use-open-weight/
- MACCUS, Tencent says Hy4 ranks among the best open models: https://maccus.fr/en/tencent-hy4-ai-model/
- ShiftDelete, Tencent Hy4: https://shiftdelete.net/en/tencent-hy4
- Reuters, Tencent rolls out AI model trained with limited cybersecurity data: https://943jackfm.com/2026/11/04/tencent-rolls-out-ai-model-trained-with-limited-cybersecurity-data-amid-hacking-fears/