Tencent's Hy4 Brings 1M Context and Agent Tools at a Tenth the Cost

Tencent’s Hunyuan team has released Hy4 , the full follow-up to its August preview. The 770-billion-parameter mixture-of-experts model adds native image understanding and computer control through graphical interfaces and terminals while retaining a 1-million-token context window.

Hy4 scores 35 on the Artificial Analysis Intelligence Index v5, a composite evaluation of reasoning, knowledge and coding. That ties Grok 4.7 (xhigh), trails Claude Haiku 5.5 (max) by one point and places Hy4 behind only Kimi K3.1 Pro, at 46, among models from Chinese labs.

For developers, Hy4 combines competitive agent benchmarks with API prices far below several similarly ranked models. Its main tradeoffs are high output volume, an 89% hallucination rate on one third-party test and the absence of full-release weights for self-hosting.

More training, new modalities

The August preview already used a 770-billion-parameter mixture-of-experts architecture with 49 billion parameters active for each token. This design routes each token through part of the network, reducing the computation required compared with activating every parameter.

The full release retains those parameter counts and the 1-million-token context window. Tencent says it expanded pre-training, which builds the model’s base capabilities, and post-training, which tunes reasoning, tool use and response behavior. Vision and computer-operation capabilities are new to this release.

Hy4 represents a substantial increase from the Hy3 series, which had fewer than 300 billion parameters. Tencent says the Hunyuan team has shipped a major version about every two months since rebuilding its training infrastructure in February.

Agents lead the benchmark profile

Hy4’s strongest third-party results cluster around agentic work, where a model plans multiple steps, edits files, calls tools and runs commands with limited supervision.

The 1-million-token context window can hold large repositories or extensive document collections in a single request. Developers still need workload-specific tests for retrieval accuracy, instruction retention and performance near the context limit.

Across the Intelligence Index evaluations, Hy4 generated 160 million output tokens, more than twice the 75 million median. Longer reasoning traces can increase latency and cost, especially because output tokens carry a higher price than input tokens.

Hallucination complicates the knowledge score

AA-Omniscience assigns Hy4 an 89% hallucination rate, compared with 37% for Kimi K3.1 Pro. The benchmark reports accuracy and hallucination as separate measures, so the percentages are not complements. Hy4 can answer many questions correctly while producing unsupported claims when it lacks reliable information.

Tencent’s internal evaluation reports that the full release reduced hallucination from 12.95% to 7.2% and guessing errors from 23.1% to 5.6%. Different datasets, prompts and scoring methods can produce large gaps between internal and third-party results, and the published figures do not provide a matched comparison.

Vision reaches the desktop

Native vision allows Hy4 to inspect detailed images, analyze charts and solve mathematics problems that include visual material. Its computer-use layer can interact with graphical interfaces and command lines, while tool calling lets an agent connect those actions across a multi-step workflow.

Tencent identifies office collaboration, financial analysis, front-end development and game development as target workloads. The combination supports tasks such as reading a chart, updating code, running a test and checking the result in an application.

Deployments that grant terminal or graphical access require sandboxing, restricted credentials, tool allowlists and confirmation before destructive actions. Those controls matter because an incorrect command can affect files, services or external accounts.

Low prices, no full checkpoint

Tencent lists three API rates for Hy4, with output tokens costing about three times as much as uncached input.

At an Intelligence Index score of 35, Artificial Analysis lists Hy4 at roughly one-tenth the cost of Grok 4.7 (xhigh). It costs about one-thirteenth as much as Qwen3.8 Max (0902), which scores 34, and about twice as much as Claude Haiku 5.5 (max), which scores 36. Actual spending depends on the mix of input, cached input and output tokens.

Access spans Tencent products, its cloud platform and third-party gateways:

• Tencent applications: WorkBuddy/CodeBuddy, Yuanbao, the WeCom smart assistant and ima.
• APIs and gateways: Tencent Cloud TokenHub, OpenRouter, Hermes and OpenCode.
• Modalities: Text and image input with text output.
• Context: Up to 1 million tokens.
• Planned variant: A cheaper Hy4 Flash model that retains vision and agent capabilities.

The August preview’s Apache 2.0 weights remain available on Hugging Face. Tencent has not published weights for the full release, leaving self-hosted deployments on the preview checkpoint for now.

Best use cases and hard limits

Hy4’s profile suits coding agents, document-heavy office automation, financial analysis, front-end work and game-development tools that use long context, images and commands. Teams evaluating it should run representative tasks against Claude Haiku 5.5 and Kimi K3.1 Pro, including tests for latency, tool reliability and total output volume.

Factual question-answering requires retrieval grounding, citations or a separate verification layer because of the Omniscience hallucination result. Computer-use deployments need restricted execution environments, while organizations that require local weights must continue using the preview or wait for Tencent to publish the full checkpoint.