Tencent Hy4 Lands in OpenClaw: 1M Context, Image Understanding, Long-Horizon Agent Gains

OpenClaw now supports Tencent Hy4, per a post on r/openclaw. The integration ships with image understanding, a 1M-token context window, and what the source describes as improvements to coding and long-horizon agent tasks.

What ships with the integration

• Image understanding — Hy4 accepts image input through OpenClaw, so agent flows that mix screenshots, diagrams, or UI captures with text no longer need a separate vision model hop.
• 1M-token context window — the headline number. For OpenClaw users running agent loops over large repos, log dumps, or multi-file refactors, a 1M window means fewer truncation hacks and fewer mid-task context resets.
• Coding and long-horizon agent improvements — the source explicitly calls out gains in coding and long-horizon agent tasks. Long-horizon here is the relevant part: multi-step agents that chain tool calls over many turns are where context decay usually shows up first, and Tencent is positioning Hy4 as better there than prior iterations.

Where to get it

The source lists three routing options: Tencent TokenHub, TokenPlan, and OpenRouter. If you're already running OpenClaw against an OpenRouter endpoint, that's the lowest-friction path — swap the model string and keep the rest of your config. TokenHub and TokenPlan are Tencent's own access routes; use those if you want the model directly from the provider rather than through a third-party router.

Practical note on the 1M context: larger windows don't make attention free. Cost scales with tokens sent, and quality on retrieval-style tasks can degrade in the middle of very long contexts. Benchmark Hy4 on your own long-horizon agent traces before moving production workloads over — especially the coding loops, where a model that holds context but drifts on tool-call syntax burns more time than it saves.

The source post doesn't include benchmark numbers against competing models, so treat the coding and agent claims as vendor-stated until you've run your own eval. If you're already on OpenClaw, the upgrade is a model config change; if you're deciding whether to switch, the 1M window and image input are the concrete deltas worth testing.