{"slug": "tencent-hy4-open-weights-run-a-770b-model-with-computer-use", "title": "Tencent Hy4 Open Weights: Run a 770B Model With Computer Use", "summary": "Tencent released the full Hy4 on November 4, 2026 under Apache 2.0, a 770B-parameter Mixture-of-Experts model with a native one-million-token context window that adds vision, GUI automation and CLI control to the text-only August preview. Tencent reports the hallucination rate fell from 15.76% to 5.45% on internal search evaluations and the speculative error rate from 23.1% to 5.6%, while API pricing stayed at $0.834 per million input tokens and $2.501 per million output tokens on Tencent Cloud, with a 50% TokenHub discount through December 31. Only 49 billion of the 770 billion parameters activate per token, and self-hosting requires 16 B200s or 8 B300s because the FP8 weights alone are about 924 GB, making the API the practical path for teams without a multi-node GPU cluster.", "body_md": "Tencent shipped the full Hy4 today, and the headline is not “big model from China.” It is the first frontier-class open-weight model that can use a computer — licensed under Apache 2.0. The August preview was text-only. Today’s full release adds vision, GUI automation, and CLI control in the same package. That combination — 770B parameters, one million token context, native computer use, fully open weights — does not exist anywhere else right now.\n\n## Why Computer Use on an Open Model Changes Things\n\nClaude Opus 5 and GPT-5 both have computer use. Neither can be self-hosted. Neither can be fine-tuned for your specific toolchain. With Hy4 under Apache 2.0, you can build a computer-use agent that runs on your own infrastructure, train it on your applications, and deploy it without hitting rate limits or platform restrictions. Tencent specifically trained Hy4 on Blender, CAD software, and long-horizon workflows that span multiple applications — connecting visual perception with planning and tool execution.\n\nOpen computer use agents were a theoretical possibility before today. Now there is a 770B model you can download, fine-tune, and ship.\n\n## What the Full Launch Actually Added\n\nThe August preview covered text generation and reasoning. The November 4 full release extends it in four directions:\n\n- **Vision:** chart interpretation, document analysis, reasoning across multiple images\n- **Computer use:** GUI and CLI operation, trained on professional applications including Blender and CAD\n- **Accuracy:** hallucination rate dropped from 15.76% to 5.45% on internal search evaluations; speculative error rate fell from 23.1% to 5.6%\n- **Reliability:** gains in instruction following, long-context understanding, and deep research tasks\n\nAPI pricing stayed flat from preview: $0.834 per million input tokens, $2.501 per million output tokens on Tencent Cloud, with a 50% discount on TokenHub through December 31.\n\n## The Architecture: Why 770B Does Not Mean 770B Per Call\n\nHy4 is a Mixture-of-Experts model. Total parameters are 770 billion, but only 49 billion activate per token — roughly the same compute footprint as a 49B dense model on each inference call. The backbone is 78 layers: one standard dense layer followed by 77 MoE layers, each with 256 routed experts and one shared expert. It also ships with a built-in speculative decoding layer (10B total, 0.7B active) that accelerates output generation.\n\nThe 1M-token context window is native, not a retrieval trick. For long document workflows or codebases that do not fit in standard context windows, that matters.\n\n## How to Run Tencent Hy4\n\nThe weights are live on [Hugging Face](https://huggingface.co/tencent/Hy4-preview) (`tencent/Hy4-preview`), ModelScope, GitCode, and CNB — both BF16 and FP8 quantized variants. [The launch was announced](https://siliconangle.com/2026/11/04/tencent-releases-hy4-powerful-frontier-llm-introducing-visual-understanding-and-computer-use/) on November 4, 2026. Apache 2.0 means commercial use and fine-tuning are permitted without restriction.\n\n**Start with the API.** OpenRouter has it available today. Tencent Cloud is the direct option if you are already in that ecosystem.\n\n**Self-hosting hardware reality check:** The FP8 quantized weights alone are approximately 924 GB. Eight H100 80GB GPUs give you 640 GB total — not enough. The [official deployment floor](https://technode.com/2026/11/04/tencent-launches-hy4-with-vision-and-computer-use-capabilities/) is 16 B200s or 8 B300s. If you do not have a multi-node GPU cluster, the API is your practical path for now.\n\nFor teams that do have the hardware, the two supported serving stacks are vLLM and SGLang:\n\n```\n# Download the weights\nhuggingface-cli download tencent/Hy4-preview --local-dir ./hy4-preview\n\n# Serve with vLLM (v0.29.0+)\nvllm serve tencent/Hy4-preview\n# SGLang via Docker\ndocker run --gpus all --shm-size 32g -p 30000:30000   -v ~/.cache/huggingface:/root/.cache/huggingface   --env \"HF_TOKEN=<secret>\" --ipc=host   lmsysorg/sglang:hy4-preview   python3 -m sglang.launch_server --model-path \"tencent/Hy4-preview\"\n```\n\nBoth stacks expose an OpenAI-compatible API endpoint. Fine-tuning pipelines are supported through DeepSpeed, LLaMA-Factory, and ms-swift. The [SGLang deployment docs](https://docs.sglang.io/cookbook/autoregressive/Tencent/Hy4-Preview) cover the full configuration options.\n\n## Benchmarks: Honest Assessment\n\nOn SWE-bench Multilingual, Hy4 scores 82.9 — ahead of GLM-5.3 (81.3) and Kimi K3.1 Pro (80.8). An internal expert evaluation across 203 engineering tasks (163 evaluators, scored out of 4.0) put Hy4 at 2.99, with a 51.2% win rate against Kimi K3 and 46.8% against GLM-5.3.\n\nThe honest weakness: SWE-Marathon, which tests sustained coding over long sessions. Hy4 scores 31.9 there — GLM-5.3 scores 42.5 and Kimi K3 scores 42.0. On Humanity’s Last Exam, Hy4 lands at 43.4 versus Claude Opus 5’s 53.2. Tencent does not hide this. The company explicitly acknowledged that Hy4 still trails Claude and GPT on some evaluations — a more credible posture than the usual “SOTA on everything” claims.\n\n## Who Should Care\n\nIf you are building computer-use automation and need to stay off closed APIs — for cost, latency, data privacy, or the ability to fine-tune — Hy4 is the only serious open option today. The SGLang Docker recipe gets you running in an afternoon if you have the hardware. OpenRouter gives you API access right now while you evaluate whether self-hosting is worth it.\n\nTencent also mentioned a Hy4 Flash variant targeting significantly lower cost and faster inference. No timeline given, but it is worth watching for teams that need high-throughput deployments at scale.", "url": "https://wpnews.pro/news/tencent-hy4-open-weights-run-a-770b-model-with-computer-use", "canonical_source": "https://byteiota.com/tencent-hy4-open-weights-770b-computer-use/", "published_at": "2026-11-04 20:06:05+00:00", "updated_at": "2026-11-04 20:23:42.533325+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-tools", "ai-infrastructure", "artificial-intelligence"], "entities": ["Tencent", "Hy4", "Hugging Face", "ModelScope", "GitCode", "CNB", "Tencent Cloud", "OpenRouter"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/tencent-hy4-open-weights-run-a-770b-model-with-computer-use", "markdown": "https://wpnews.pro/news/tencent-hy4-open-weights-run-a-770b-model-with-computer-use.md", "text": "https://wpnews.pro/news/tencent-hy4-open-weights-run-a-770b-model-with-computer-use.txt", "jsonld": "https://wpnews.pro/news/tencent-hy4-open-weights-run-a-770b-model-with-computer-use.jsonld"}}