ai-info

AI news

Tencent Hy4 Is Live on OpenRouter at Half the Preview's Price

·

Tencent Hy4 Is Live on OpenRouter at Half the Preview's Price

Short answer

  • Tencent released the final Hy4 today (Nov 4): 770B MoE / 49B active / 1M context, first Hunyuan generation with vision understanding and Computer Use. It is already live on OpenRouter as tencent/hy4 — we pulled the listing and verified $0.417/M input, $1.2505/M output, 1,048,576 context.
  • The odd part: tencent/hy4-preview is still listed at $0.834/$2.501 — exactly 2× the final model, same context window. Official list pricing is unchanged at ¥6/¥18 per 1M tokens (cache hit ¥0.3), but on OpenRouter, switching the model ID to hy4 halves your bill today with zero code changes beyond the string.
  • There is also a time-boxed discount: 50% off list on Tencent Cloud TokenHub until Dec 31, 2026 — that puts the full 770B model at roughly ¥3/¥9, which is the cheapest way to run a frontier-tier open-weights-class model right now. Weights themselves ship "soon" with a model system card; self-hosters should wait for the actual repos, not the press release.
Hunyuan Hy4 official list price in CNY (6/18/0.3 per 1M tokens) and OpenRouter listed prices showing the final hy4 at half the hy4-preview tag
Left: official list pricing, unchanged from preview. Right: what OpenRouter actually lists today — the final model at half the preview tag, same 1M context.

Tencent's own framing is "same price, much more model": Hy4 preview came out on Aug 28, and the final version extends pretraining and post-training, adds vision (fine-grained image reading, chart analysis, visual math) and — for the first time in the Hy line — Computer Use, with dedicated training on Blender, CAD, and long-horizon cross-app tasks. Benchmarks cited by Tencent put it in the open-source first tier: in a 213-expactor, 579-task finance blind test it scored 3.04 vs 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for Kimi K3.1 Pro and GLM-5.3, and its self-reported hallucination rate dropped from 12.95% to 7.2%. Treat those as vendor numbers — the pricing and the OpenRouter listing are the parts we verified directly.

For developers outside China, the practical read:

API users: the model is reachable now via OpenRouter (plus Hermes and OpenCode per Tencent), so no China-facing billing setup needed. If you were on tencent/hy4-preview through OpenRouter, move to tencent/hy4 today — the outputs are the upgraded model and the listed price is half. If you can bill Tencent Cloud directly, the TokenHub 50% promo runs through end of year; that's ¥3 in / ¥9 out per 1M tokens, and the cache-hit rate (¥0.3/M, ¥0.15 under the promo) is what matters if your agent workflow re-sends long stable prefixes — the same mechanics we broke down in our caching guide.

Self-hosters: hold. Tencent says weights and a system card come "soon", in the same pattern as Hy4 preview's rollout. When they land, at 49B active parameters this is a plausibly 4×H200-class deployment — meaningfully lighter than DeepSeek-V4-Pro but heavier than anything consumer hardware touches. We'll write the self-host decision the day the repos are actually up, same as we did for Kimi's weight drop.

On the safety trimming: Tencent says Hy4's training kept only defensive security data (vulnerability fixing, PoC generation) and stripped offensive content, with audits by its Xuanwu and Keen labs showing no standout offensive capability. That's the same trade Moonshot made with K3.1 Pro's open weights — if your use case is offensive security research, both of China's new open-weights flagships are now deliberately blunted, and closed models remain the only option there.

One more date to circle: Hy4 flash is also "coming soon", pitched as the cheap tier — given that the full 770B model already costs $1.25/M output on OpenRouter, the flash variant will be competing directly with DeepSeek's flash pricing. That comparison is the obvious follow-up piece.

Verified directly on 2026-11-04: OpenRouter listing for tencent/hy4 and tencent/hy4-preview (prices, context length), via the public models API. Official list pricing (¥6/¥18/¥0.3), the TokenHub 50% promo through 2026-12-31, benchmark scores, hallucination figures, and the "weights soon" commitment are per Tencent's announcement as reported by multiple outlets (Sohu/Qianjiang, Tencent News, NetEase Finance) — vendor-reported, not independently audited.

Sources: OpenRouter models API (tencent/hy4 listing, checked 2026-11-04), Sohu / Qianjiang Evening News — Tencent Hunyuan releases Hy4, Tencent News — Hy4 announcement details