Tencent releases Hy4 with a 1M-token context window in OpenClaw

• Hy4 brings vision and a 1M-token context window to OpenClaw, according to the integration announcement.
• Hosted access is live, with weights and Hy4 Flash planned per AiBattle_; TencentHunyuan says Chutes support is coming.
• Tencent excluded offensive cybersecurity training material while retaining defensive examples, according to the excerpt shared by AiBattle_.

Cache pricing steals the show in Plotly's independent test: 91% of Hy4's 22.7 million input tokens came from cache. Tencent's launch demos include locating an approximately two-meter rooftop component in a 2.8-gigapixel aerial image.

TencentHunyuan confirmed Hy4 support in OpenClaw, whose announcement lists three access routes:

• Tencent TokenHub
• Tencent TokenPlan
• OpenRouter

OpenRouter exposes the model as tencent/hy4. Tencent says its standard API prices are unchanged from Hy4 Preview, while OpenRouter lists a 50% launch discount.

| Per million tokens | Tencent standard API | OpenRouter launch pricing |
| --- | ---: | ---: |
| Input | $0.834 | $0.417 |
| Output | $2.501 | $1.251 |
| Cached input | $0.042 | $0.021 |

Vision and computer use are new relative to Hy4 Preview, according to Tencent's technical release. The capabilities and training targets include:

• Fine-grained image interpretation and complex chart understanding.
• Computer actions across command-line and graphical interfaces.
• Targeted Blender and CAD training, including long-horizon workflows spanning multiple applications.
• Adaptation for Unreal Engine, Unity, and Godot, extending into multiplayer game development.

Tencent worked with professional developers from its Interactive Entertainment Group to build game-development training data and evaluations, focusing on implementation, gameplay logic, state tracking, and asset pipelines.

WorkBuddy's WeCom integration connects Hy4 to five sources of work context:

1. Chat history
2. Email
3. Documents
4. Calendars
5. Meeting notes

Tencent says the assistant can use that context to track projects and prepare meeting briefs. Team-collaboration features are scheduled to arrive with WorkBuddy's new version of Projects.

Tencent's reported improvements combine first-party evaluations with a finance partner's benchmark:

• GameDev: +19.1% versus Hy4 Preview.
• GameLogic: +9.8% versus Hy4 Preview.
• Internal hallucination rate: 15.76% → 5.45%, -10.31 percentage points.
• RabyteTech's IRAB: 61.63 across 101 investment-research tasks, +12.03 points versus Hy4 Preview.

Tencent also acknowledges that leading proprietary models remain ahead on long-horizon coding and research-level mathematics.

Hy4 passed 42 of 43 questions for $2.11 in Plotly's data-analytics evaluation. Plotly used its Studio Embedded analytics harness at OpenRouter's default settings, querying a synthetic wind-energy data warehouse.

The six categories broke down as follows:

• Analytics: 20/20
• Ambiguous questions: 5/5
• Impossible requests: 5/5
• Data quality: 3/3
• Statistics traps: 5/5
• Hidden patterns: 4/5

GPT-6.1 Sol scored 43/43 for $2.61. Each model received one run; Plotly says repeat runs of other models have varied by up to three points.

Median time per question was 117 seconds, compared with 18 seconds for GPT-6.1 Sol and 38 seconds for Grok 4.7. One Hy4 query paired every row of one large table with every row of another and ran for almost an hour; the harness imposed no query timeout.

On its missed question, Hy4 identified a genuine excess failure rate in one component type but never examined the part of the data containing the hidden pattern. It also ruled out a cause it had not checked.

Tencent describes a training-data filter that AiBattle_ highlighted:

Model weights and system cards are due “in the coming days,” according to Tencent's November 4 launch announcement. Tencent names vLLM and SGLang as ecosystem partners for the weight release.

Hy4 Flash is planned as a faster, more cost-efficient follow-up, with closer product co-design.

Chutes support is also forthcoming: