{"slug": "tencent-hy4-770b-vision-computer-use-model-developer-guide", "title": "Tencent Hy4: 770B Vision + Computer-Use Model, Developer Guide", "summary": "Tencent released Hy4 on November 4, a 770-billion-parameter mixture-of-experts model that adds vision and computer-use capabilities and is priced at roughly $0.42 per million input tokens on OpenRouter, about ten times cheaper than Claude Opus 5.5. Tencent-reported benchmarks show Hy4 scoring 78.9 on OSWorld 2.1 versus Claude Opus 5.5's 81.8, but 30.0 on ProgramBench versus 56.5 and 13.0 on Frontier Math versus GPT-6.1 Sol's 97.5, and third-party hallucination testing found an 89% rate against Tencent's claimed 5.45%. Despite calling Hy4 open source, Tencent's weights were not publicly available as of November 6 — the Hugging Face repository returned a 401 error and the GitHub repo a 404 — with Tencent saying weights and a system card are coming \"in the coming days.", "body_md": "Tencent dropped Hy4 on November 4 — a 770-billion-parameter mixture-of-experts model that now sees images and operates computers. At roughly $0.42 per million input tokens on OpenRouter, it undercuts Claude Opus 5.5 by around ten times. The marketing says “open source” and “#1 open-weight model.” The reality is more nuanced, and developers deserve the unvarnished version.\n\n## Computer Use Is the Headline Feature\n\nThe August 2026 Hy4 preview was text-only. This release adds something different: the model reads on-screen text, clicks UI elements, runs CLI commands, and chains tasks across multiple applications. Tencent demonstrated booking appointments, extracting spreadsheet data, and managing email workflows. For agentic work, [Hy4 hit 78.9 on OSWorld 2.1](https://www.unite.ai/tencent-releases-full-hy4-model-with-vision-and-computer-use/) — Claude Opus 5.5 leads at 81.8, but that gap is small enough to matter when you factor in cost.\n\nThe model was specifically post-trained for professional tooling: Blender, CAD applications, Unreal Engine, Unity, and Godot. If you’re building agents that need to operate real software rather than just generate code, that targeted training is worth noting.\n\n## How to Access Hy4 Today\n\nYou don’t need to wait for weights. Hy4 is live right now across several platforms:\n\n- **OpenRouter** — model ID`tencent/hy4` , available immediately\n- **CodeBuddy and WorkBuddy** — free through November 15, 2026\n- **Tencent Cloud TokenHub** — 50% off through December 31, 2026\n- **Hermes and OpenCode** — rolling out now\n\nThe [OpenRouter integration](https://openrouter.ai/tencent/hy4) means you can drop it into any OpenAI-compatible client. Change the base URL to `https://openrouter.ai/api/v1` and set the model to `tencent/hy4` — that’s the full migration.\n\n## The Price Math (With a Catch)\n\nList price is $0.834/$2.501 per million input/output tokens. OpenRouter’s promotional rate brings that down to roughly $0.42/$1.25 — about ten times cheaper than Grok 4.7 and thirteen times cheaper than Qwen3.8 Max at comparable capability levels.\n\nThe catch: Hy4 is verbose. Third-party analysis found it averaging 160 million output tokens against a 75 million median for similar models. That doubling of output volume can quietly eat the per-token savings. If you’re running high-volume pipelines, benchmark your actual token usage before committing.\n\n## Benchmarks: The Honest Picture\n\nAll twelve benchmarks in Tencent’s launch materials are Tencent-reported. No independent verification exists yet for the full November model. Keep that in mind:\n\n- **Finance workflows:** Ranked first among open-weight models in an evaluation with 213 experts on 579 real-world tasks (3.04 vs. DeepSeek-V4.1-Flash’s 2.97)\n- **SWE-bench Multilingual:** 82.9% resolved — a solid result for multi-language code repair\n- **Coding agents:** 30.0 on ProgramBench versus Claude Opus 5.5’s 56.5 — a meaningful gap\n- **Terminal automation:** 44.4 on Terminal Bench 4.0 versus Kimi K3.1 Pro’s 66.8 — not ideal for pure CLI work\n- **Math:** 13.0 on Frontier Math versus GPT-6.1 Sol’s 97.5 — do not use this for quantitative reasoning\n\nThird-party hallucination testing also surfaced an 89% rate — a stark contrast to Tencent’s claimed 5.45%. [Different evaluation sets likely explain the discrepancy](https://www.digitalapplied.com/blog/tencent-hy4-release-what-is-confirmed), but the variance is wide enough to treat both numbers as directional, not authoritative.\n\n## “Open Source” Is Still a Pledge\n\nThis needs saying plainly: Tencent is calling Hy4 open source, but the model weights are not publicly available as of November 6. The Hugging Face repository for `tencent/hy4` returns a 401 error. The GitHub repo returns a 404. The August 2026 preview shipped with Apache 2.0, but that license does not automatically extend to the full release. Tencent says weights and a system card are coming “in the coming days.” Until then, you’re using a hosted API, not open-source software.\n\n## When to Use Hy4 vs. the Alternatives\n\nHy4 earns its place for finance and document workflows, long-context tasks that fill the 1M-token window, and computer-use agents where cost pressure is real. For coding accuracy and terminal automation, Claude Opus 5.5 and Kimi K3.1 Pro still lead on the available numbers. [Once weights drop, the self-hosting story changes](https://alphasignal.ai/news/tencent-s-hy4-brings-1m-context-and-agent-tools-at-a-tenth-the-cost) — Hy4 becomes a compelling option for teams with data-residency requirements that can’t use hosted APIs.\n\nUse it free in CodeBuddy through November 15 and form your own opinion before the promotional window closes.", "url": "https://wpnews.pro/news/tencent-hy4-770b-vision-computer-use-model-developer-guide", "canonical_source": "https://byteiota.com/tencent-hy4-770b-vision-computer-use/", "published_at": "2026-11-06 10:06:19+00:00", "updated_at": "2026-11-06 10:23:02.584301+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-products", "ai-tools", "artificial-intelligence"], "entities": ["Tencent", "Hy4", "OpenRouter", "Claude Opus 5.5", "DeepSeek-V4.1-Flash", "Kimi K3.1 Pro", "GPT-6.1 Sol", "Hugging Face"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/tencent-hy4-770b-vision-computer-use-model-developer-guide", "markdown": "https://wpnews.pro/news/tencent-hy4-770b-vision-computer-use-model-developer-guide.md", "text": "https://wpnews.pro/news/tencent-hy4-770b-vision-computer-use-model-developer-guide.txt", "jsonld": "https://wpnews.pro/news/tencent-hy4-770b-vision-computer-use-model-developer-guide.jsonld"}}