Tencent Hy4: A Developer's Guide to the 770B Model

Tencent Hy4 is a 770-billion-parameter mixture-of-experts language model released on 4 November 2026. It is the first model in Tencent's Hunyuan line that can both see images and operate a computer, reading screens, clicking interface elements and running commands across applications. It is available now through an API, with its weights promised but not yet published.

Key takeaways

• Hy4 is a 770B mixture-of-experts model with 49B active parameters per token and a 1-million-token context window.

• Its headline feature is computer use: GUI and CLI operation, with targeted training for Blender, CAD, Unreal Engine, Unity and Godot.

• It is live on OpenRouter as tencent/hy4 at a list price of $0.834/$2.501 per million input/output tokens, with a promotional rate near $0.42/$1.25.

• Every benchmark in the launch materials is Tencent-reported. No independent verification exists yet for the full November model.

• The weights are not public as of 6 November 2026, so calling Hy4 open source is a promise, not a fact you can download.

What is Tencent Hy4?

Hy4 is the fourth generation of Tencent's Hunyuan foundation model line, and the successor to Hy3. A preview shipped on 28 August 2026 as a text-only model with its weights published under Apache 2.0. The full release on 4 November 2026 adds the two capabilities the preview lacked: vision understanding and computer use.

The architecture is a mixture-of-experts design: 770 billion total parameters, with 49 billion active per token, spread across 78 layers and 256 routed experts plus one shared expert. The context window is 1,048,576 tokens, which puts it in the same long-context class as the models it competes with. A 10-billion-parameter multi-token prediction layer sits on top, which is why some listings show 780B parameters instead of 770B. Both numbers describe the same model.

Tencent positions Hy4 around five kinds of work rather than a single headline benchmark: coding, office tasks, data analysis, game development and scientific research. That framing matters, because the model's strengths and weaknesses split cleanly along those lines.

What can Hy4 actually do?

Computer use is the feature that separates Hy4 from the August preview and from most open-weight models. Tencent says the model reads on-screen text, clicks UI elements, runs CLI commands and chains tasks across multiple applications. The demonstrations cover booking appointments, extracting spreadsheet data and managing email workflows.

The post-training is unusually specific. Hy4 was trained for professional software including Blender and CAD applications, plus game engines: Unreal Engine, Unity and Godot. If you are building agents that must operate real software rather than just generate code, that targeted training is the most interesting part of the release. Tencent also says the model was trained on long-horizon workflows that span several applications, so an agent can plan a multi-step task without losing the thread.

Vision is the second addition. Hy4 processes images, analyses complex charts and performs visual mathematical reasoning. That makes it usable for document-heavy workflows: reading a scanned invoice, summarising a dashboard screenshot, or extracting values from a chart.

On agentic work, Hy4 scored 78.9 on OSWorld 2.1 in Tencent's own testing, against 81.8 for Claude Opus 5.5. That is a small gap for a model that costs roughly a tenth as much on the promotional rate, and it is the strongest argument for trying Hy4 in a computer-use agent.

How do you try Hy4 today?

You do not need to wait for weights. Hy4 is live now on OpenRouter as tencent/hy4, on Tencent Cloud TokenHub, and inside Tencent's own CodeBuddy and WorkBuddy products, which are free from 4 to 15 November 2026. TokenHub runs at 50% off through 31 December 2026. Hermes and OpenCode are rolling it out as well.

The OpenRouter route is the easiest for developers, because the API is OpenAI-compatible. Point your existing client at a new base URL and change the model name, and nothing else needs to change:

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="YOUR_OPENROUTER_API_KEY",
)

response = client.chat.completions.create(
    model="tencent/hy4",
    messages=[
        {"role": "system", "content": "You are a coding assistant. Answer concisely."},
        {"role": "user", "content": "Write a Python function that retries a failed HTTP request with exponential backoff."},
    ],
)

print(response.choices[0].message.content)

Vision input works through the same client, by passing an image alongside the text:

response = client.chat.completions.create(
    model="tencent/hy4",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Read the values in this chart and summarise the trend in two sentences."},
            {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
        ],
    }],
)

On OpenRouter, the model has shown a median throughput near 63 tokens per second, a median time to first token near 2.9 seconds, and 99.87% availability over its first three days. Those are aggregator-reported figures, not a formal benchmark, but they suggest the hosted endpoint is stable enough to build against.

What do the benchmarks actually say?

All twelve benchmarks in Tencent's launch materials are Tencent-reported. Tencent reran rival models under consistent settings, so the comparisons are fairer than quoting each vendor's own marketing, but they are still not an independent leaderboard. No independent organisation has rated the full November model yet. Read the numbers with that label attached.

The strong results first. Hy4 resolved 82.9% of tasks on SWE-bench Multilingual, a solid result for multi-language code repair. In a finance evaluation run with 213 finance professionals across 579 real-world tasks, Hy4 scored 3.04 out of 4.00 and ranked first among open-weight models, ahead of DeepSeek V4.1 Flash at 2.97 and Kimi K3.1 Pro at 2.68. On E-Bench 2.0 it scored 70.0, second of seven models tested. Tencent also reports that hallucination fell from 15.76% in the preview to 5.45% in the full model, and that game development output improved 19.1% over the preview.

The weak results are just as informative. On ProgramBench, a coding-agent benchmark, Hy4 scored 30.0 against 56.5 for Claude Opus 5.5. On Terminal Bench 4.0 it scored 44.4 against 66.8 for Kimi K3.1 Pro, so pure CLI automation is not its strength. On FrontierMath it scored 13.0 against 97.5 for GPT-6.1 Sol, which means quantitative reasoning is a clear weakness. Do not use this model for maths.

One number deserves a warning of its own. A third-party hallucination test reported an 89% hallucination rate for Hy4, against Tencent's claimed 5.45%. The two tests almost certainly used different evaluation sets, so neither number is authoritative on its own. The gap is wide enough that you should treat both as directional and run your own checks on your own data before trusting the model in production.

How much does Hy4 cost?

The list price is $0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042. OpenRouter's promotional rate brings that to roughly $0.42/$1.25, about ten times cheaper than Claude Opus 5.5 at comparable capability levels on the tasks where Hy4 is competitive.

There is a catch, and it is one that shows up in the bill rather than the price list. Third-party analysis found Hy4 averaging about 160 million output tokens against a 75 million median for comparable models. Hy4 is verbose. If your workload generates a lot of output, the extra tokens can quietly eat the per-token savings. Benchmark your actual token usage on your own prompts before committing a high-volume pipeline to it.

Is Hy4 really open source?

This needs saying plainly. Tencent is calling Hy4 open source, but the model weights are not publicly available as of 6 November 2026. The Hugging Face repository for tencent/hy4 returns a 401 error, and the GitHub repository returns a 404. Tencent says the weights and a system card are coming in the coming days. Until then, you are using a hosted API, not open-source software.

The August preview did ship with Apache 2.0 weights, but that licence does not automatically extend to the full release. Tencent has partnered with vLLM and SGLang for serving support, which suggests the release is genuinely planned, but a partnership announcement is not a download link.

When the weights do land, self-hosting will not be a weekend project. The reference configuration assumes 16 B200 GPUs, or 8 B300s. That is a data centre, not a workstation. The self-hosting story matters most for teams with data-residency requirements that cannot use hosted APIs, and for them the current position is simple: wait for the weights, and do not plan a migration around a release date that has not been announced.

Should you use Hy4 for your next agent?

Hy4 earns its place for finance and document workflows, for long-context tasks that fill the 1-million-token window, and for computer-use agents where cost pressure is real. The OSWorld result and the finance evaluation both point the same way: it is close enough to the frontier on agentic work, at a fraction of the price.

For coding accuracy and terminal automation, Claude Opus 5.5 and Kimi K3.1 Pro still lead on the available numbers. For quantitative reasoning, Hy4 is simply the wrong tool. And for anything where the output must be terse or the bill must be predictable, measure the verbosity first.

The honest summary is that Hy4 is a strong, cheap, genuinely useful model with a marketing label that currently overstates its openness. Use it free in CodeBuddy through 15 November and form your own opinion before the promotional window closes. Watch for the weights, verify the benchmarks independently when they appear, and do not build a self-hosting plan on a promise.

Originally published on LogicStack. Run the code examples right in your browser with our free interactive Python and SQL playgrounds — no setup, no sign-up.