In-depth review of Tencent Hunyuan Hy4: Official 1M-context version launched, Agent capabilities significantly enhanced, $0.42 per million tokens

The open-weight version released at the end of August still carried the “Preview” suffix; on November 4, Tencent Hunyuan released the official Hy4 versionas an API. This time it’s not just publishing a technical report and calling it a day—Tencent’s official provider is already live on OpenRouter, running at 100% uptime, priced at $0.417 per million input tokens and $1.25 per million output tokens, roughly 40% of OpenAI GPT-6.1 Terra’s price, with full 1M context support.

For those of us who run Agents on models daily, there’s only one critical question:Can it truly replace our current GLM 5.3 + K3.1 stack? Kimi ?This article lays out official specifications, pricing, blind-test data, and verified capability boundaries.

What is Hunyuan Hy4: From Preview to official release

Hy4 is the fourth-generation flagship from Tencent Hunyuan (Tencent Hy Team), built on a MoE sparse-activation architecture. Preview weights were open-sourced on August 27 (Apache 2.0), and the official API launched on November 4. The official description of the official version is straightforward: compared to Preview,Agent capabilities and long-horizon productivity tasks saw the largest improvements—and these have been the sole focus of the Hunyuan team over the past two months.

Known core specifications (architecture data from official model card; official weights not yet synced to Hugging Face):

• 770B total parameters / 49B activated78 layers, 77-layer MoE with 256 routed experts + 1 shared expert per layer; top-8 experts activated per token
• 1M context length; up to 65,536 tokens generated per inferenceHandles long documents, large codebases, and multi-turn Agent trajectories in a single pass
• Native multimodalityAPI supports text + image input and text-only output—screenshots, design mockups, and error interfaces can be fed directly
• Attention uses Gated DSA + IndexCacheCross-layer reuse of sparse indices—the key to maintaining cost efficiency at 1M context
• Built-in MTP speculative decoding layer(10B total parameters / 0.7B activated): official vLLM and SGLang images available
• Supports reasoning_effort, tools, and structured outputsThis configuration is designed for Agent frameworks—not for chat

Core capability: long-horizon Agent execution is the primary focus

Official training data is organized by real Tencent internal job roles: software engineers, game developers, financial analysts, security specialists. Four capability domains stand out:

• Software engineeringUnderstands, plans, debugs, and verifies long-cycle development tasks—with frontend specialization in “visual taste” and interaction quality
• Office & analyticsConsolidates messy context scattered across dozens of files into documents, tables, and PPTs; significantly improves accuracy of financial models and formulas
• Game developmentGenerates playable prototypes from a single prompt and supports iterative multi-turn refinement with game engines

More telling is that blind testBlind test163 internal Tencent experts, 203 real engineering tasks, double-blind comparison. Hy4 Preview scores 2.99, GLM 5.3 scores 2.92 (wins 46.8% / ties 12.8% / loses 40.4%).Kimi K3 scores 2.94 (wins 51.2% / ties 7.9% / loses 40.9%).The lead is narrow.Yet this result was achieved against rivals all from the open-source top tier—and the official release further strengthens Agent capabilities over the Preview version, making a genuine step forward in this direction.

Pricing: $0.42 per million tokens—what does it mean?

Official release pricing (per million tokens): Input $0.417and $20 per million output tokens $1.2505,Cache read: $0.021Cache read costs only 1/20 of the input price, meaning marginal cost is sharply reduced for the most expensive usage pattern—feeding the same codebase repeatedly to an Agent.

Model	context	Input / Output (per million tokens)	Core Positioning	This Site’s Rating
Tencent Hunyuan Hy4	1M	$0.42 / $1.25	Open-source flagship Agent	8.9
GLM 5.3 FlashX	1M	$0.37 / $1.25	High-Speed Multimodal	8.7
Kimi K3.1 Pro	1M	$2 / $10	2.8T open-source flagship	8.8
GPT-6.1 Terra	1.05M	$1 / $5	Computer operation model	8.9

Horizontally, Hy4’s output price is just 1/4 that of GPT-6.1 Terra,Kimi 1/8 that of K3.1 Pro, and roughly on par with GLM 5.3 FlashX. In other words,Zhipu and Tencent are now competing head-to-head in the same pricing tier,which happens to be the primary budget range for most Chinese Agent projects.

User experience and limitations

The official team lists two known issues—both critically severe in Agent scenarios:Excessive reasoning time on complex tasks,as well asOverzealous self-verificationIn plain terms: It repeatedly checks the code it just wrote, consuming more tokens than expected and increasing latency. For token-based pricing scenarios, this means a listed price of $0.42 does not equal an actual cost of $0.42.

Two additional points to note: the official releaseopen-source weights have not yet been updated— on Hugging Face, only the August Hy4-preview and FP8 versions remain available; private deployment is limited to Preview weights; on the API side, Tencent’s official provider uses fp4-quantized inference, which differs in performance from BF16 self-hosted deployments.

Tencent Hunyuan Hy4 vs. other comparable models

Model	Parameter count	Core advantages	Weaknesses	This Site’s Rating
Hunyuan Hy4	770B / 49B activated	1M context + image input + extremely low caching cost	Overly lengthy reasoning, delayed open-source weights	8.9
GLM 5.3	Unreleased	Robust million-token context inference, mature ecosystem	Average speed	8.4
Kimi K3.1 Pro	2.8T	Largest open-source model, strong long-document reasoning	Pricing 5–8× higher	8.8
Xiaomi MiMo V2.6	Unreleased	Three-tier pricing, Flash tier extremely affordable	Subpar performance on complex Agent tasks	8.1

The conclusion is clear: Hy4 holds its ground as an “open-source, 1M-context Agent model,” priced competitively against GLM-series models rather than Kimi the series. It is not the strongest, butAgent completion rate per unit costIt is currently the most competitive model in China’s domestic AI ecosystem.

Overall Score

维度	Score	evaluate
functional completeness	8.9 / 10	Supports 1M context length, image input, tool calling, and structured output.
易用性	8.6 / 10	OpenAI-compatible API, official vLLM/SGLang Docker images—integration requires virtually zero cost.
Cost-effectiveness	9.3 / 10	Priced at $0.42/$1.25 per million input/output tokens; cache read costs only $0.021—significant advantage in long-context scenarios.
中文支持	9.2 / 10	Trained natively on Chinese corpora by a domestic team; delivers stable performance on Chinese tasks and terminology understanding.
输出质量	8.7 / 10	Slightly outperforms GLM 5.3 and Kimi K3 in blind evaluations—but the margin is narrow.

Overall rating: 8.9/10

Frequently Asked Questions (FAQ)

Is Tencent Hunyuan Hy4 free?

Its API is not free—it operates on a pay-per-use basis: $0.417 per million input tokens, $1.2505 per million output tokens, and $0.021 per million cache reads. The open-source weights (Preview version) are licensed under Apache 2.0 and can be freely downloaded for self-hosting; however, the official production weights have not yet been released.

What hardware configuration is required to deploy Hy4 locally?

With a total of 77B parameters in MoE architecture, the official recommendation is FP8-weight inference using 8-GPU tensor parallelism with the vLLM Docker image vllm/vllm-openai:hy4-preview or the SGLang Docker image, lmsysorg/sglang:hy4-previewenabled with MTP speculative decoding. It cannot run on consumer-grade GPUs in a single machine—it is designed for data-center-level deployment.

Which should I choose: Hy4 or GLM 5.3?

They are priced similarly and both support 1M context. Choose Hy4 for long-horizon programming agents or when image input is required; choose GLM 5.3GLM 5.3 for mature ecosystem support, stable throughput, and to avoid pitfalls associated with new models. Their blind evaluation scores differ by only ~0.07 points—the practical experience depends more on your prompts and toolchain.

Does Hy4 support Chinese—and how well does it perform?

Yes—Chinese is one of its native languages. Its vocabulary size is 120,832, and its official training data includes extensive internal Chinese engineering documentation from Tencent, resulting in stable performance on Chinese technical terminology and code-commenting tasks.

Summary: Worth Adding to Your Model Pool

Hy4 Stable Release Is Not “Yet Another Benchmark-Chasing Model”—It Solves a Very Real Problem:Delivering 1M-Context + Agent Tool Calling + Image Input at $0.40 per Million TokensIf your project is bottlenecked by long-context costs, it’s worth running an A/B test. But don’t expect it to save you tuning effort—its two inherent flaws—overly verbose reasoning and excessive self-verification—will genuinely erode part of its price advantage.

Want to Compare Real Pricing and Scores Across More Domestic and Overseas Models? Check Our AI Model Library and Tool Comparison Engineor continue reading:GLM 5.3 FlashX benchmarking · Kimi K3.1 Pro Review · Document your initial hypothesis and metric selection with a timestamp. Experiments show this does not fully prevent influence, but at least lets you see afterward what changed in your thinking · All Reviews.