{"id":695,"date":"2026-11-05T07:06:57","date_gmt":"2026-11-04T23:06:57","guid":{"rendered":"https:\/\/aidashxp.com\/tencent-hy4-review\/"},"modified":"2026-11-05T07:07:49","modified_gmt":"2026-11-04T23:07:49","slug":"tencent-hy4-review","status":"publish","type":"post","link":"https:\/\/aidashxp.com\/en\/tencent-hy4-review\/","title":{"rendered":"In-depth review of Tencent Hunyuan Hy4: Official 1M-context version launched, Agent capabilities significantly enhanced, $0.42 per million tokens"},"content":{"rendered":"<p class=\"wp-block-paragraph\">The open-weight version released at the end of August still carried the \u201cPreview\u201d suffix; on November 4, Tencent Hunyuan released <strong>the official Hy4 version<\/strong>as an API. This time it\u2019s not just publishing a technical report and calling it a day\u2014Tencent\u2019s 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\u2019s price, with full 1M context support.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">For those of us who run Agents on models daily, there\u2019s only one critical question:<strong>Can it truly replace our current GLM 5.3 + K3.1 stack? <a href=\"https:\/\/kimi.moonshot.cn\" target=\"_blank\" rel=\"nofollow noopener\">Kimi<\/a> ?<\/strong>This article lays out official specifications, pricing, blind-test data, and verified capability boundaries.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What is Hunyuan Hy4: From Preview to official release<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">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,<strong>Agent capabilities and long-horizon productivity tasks saw the largest improvements<\/strong>\u2014and these have been the sole focus of the Hunyuan team over the past two months.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Known core specifications (architecture data from official model card; official weights not yet synced to Hugging Face):<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>770B total parameters \/ 49B activated<\/strong>78 layers, 77-layer MoE with 256 routed experts + 1 shared expert per layer; top-8 experts activated per token<\/li>\n<li><strong>1M context length; up to 65,536 tokens generated per inference<\/strong>Handles long documents, large codebases, and multi-turn Agent trajectories in a single pass<\/li>\n<li><strong>Native multimodality<\/strong>API supports text + image input and text-only output\u2014screenshots, design mockups, and error interfaces can be fed directly<\/li>\n<li><strong>Attention uses Gated DSA + IndexCache<\/strong>Cross-layer reuse of sparse indices\u2014the key to maintaining cost efficiency at 1M context<\/li>\n<li><strong>Built-in MTP speculative decoding layer<\/strong>(10B total parameters \/ 0.7B activated): official vLLM and SGLang images available<\/li>\n<li><strong>Supports reasoning_effort, tools, and structured outputs<\/strong>This configuration is designed for Agent frameworks\u2014not for chat<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">Core capability: long-horizon Agent execution is the primary focus<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Official training data is organized by real Tencent internal job roles: software engineers, game developers, financial analysts, security specialists. Four capability domains stand out:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Software engineering<\/strong>Understands, plans, debugs, and verifies long-cycle development tasks\u2014with frontend specialization in \u201cvisual taste\u201d and interaction quality<\/li>\n<li><strong>Office &amp; analytics<\/strong>Consolidates messy context scattered across dozens of files into documents, tables, and PPTs; significantly improves accuracy of financial models and formulas<\/li>\n<li><strong>Game development<\/strong>Generates playable prototypes from a single prompt and supports iterative multi-turn refinement with game engines<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">More telling is that blind test<strong>Blind test<\/strong>163 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%).<a href=\"https:\/\/kimi.moonshot.cn\" target=\"_blank\" rel=\"nofollow noopener\">Kimi<\/a> K3 scores 2.94 (wins 51.2% \/ ties 7.9% \/ loses 40.9%).<strong>The lead is narrow.<\/strong>Yet this result was achieved against rivals all from the open-source top tier\u2014and the official release further strengthens Agent capabilities over the Preview version, making a genuine step forward in this direction.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Pricing: $0.42 per million tokens\u2014what does it mean?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Official release pricing (per million tokens): Input <strong>$0.417<\/strong>and $20 per million output tokens <strong>$1.2505<\/strong>,<strong>Cache read: $0.021<\/strong>Cache read costs only 1\/20 of the input price, meaning marginal cost is sharply reduced for the most expensive usage pattern\u2014feeding the same codebase repeatedly to an Agent.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>Model<\/th><th>context<\/th><th>Input \/ Output (per million tokens)<\/th><th>Core Positioning<\/th><th>This Site\u2019s Rating<\/th><\/tr><\/thead>\n<tbody>\n<tr><td><strong>Tencent Hunyuan Hy4<\/strong><\/td><td>1M<\/td><td>$0.42 \/ $1.25<\/td><td>Open-source flagship Agent<\/td><td>8.9<\/td><\/tr>\n<tr><td><a href=\"https:\/\/aidashxp.com\/en\/glm-5-3-flashx-review\/\">GLM 5.3 FlashX<\/a><\/td><td>1M<\/td><td>$0.37 \/ $1.25<\/td><td>High-Speed Multimodal<\/td><td>8.7<\/td><\/tr>\n<tr><td><a href=\"https:\/\/aidashxp.com\/en\/kimi-k3-1-pro-review\/\">Kimi K3.1 Pro<\/a><\/td><td>1M<\/td><td>$2 \/ $10<\/td><td>2.8T open-source flagship<\/td><td>8.8<\/td><\/tr>\n<tr><td><a href=\"https:\/\/aidashxp.com\/en\/gpt-6-1-terra-review\/\">GPT-6.1 Terra<\/a><\/td><td>1.05M<\/td><td>$1 \/ $5<\/td><td>Computer operation model<\/td><td>8.9<\/td><\/tr>\n<\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Horizontally, Hy4\u2019s output price is just 1\/4 that of GPT-6.1 Terra,<a href=\"https:\/\/kimi.moonshot.cn\" target=\"_blank\" rel=\"nofollow noopener\">Kimi<\/a> 1\/8 that of K3.1 Pro, and roughly on par with GLM 5.3 FlashX. In other words,<strong>Zhipu and Tencent are now competing head-to-head in the same pricing tier,<\/strong>which happens to be the primary budget range for most Chinese Agent projects.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">User experience and limitations<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The official team lists two known issues\u2014both critically severe in Agent scenarios:<strong>Excessive reasoning time on complex tasks<\/strong>,as well as<strong>Overzealous self-verification<\/strong>In 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.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Two additional points to note: the official release<strong>open-source weights have not yet been updated<\/strong>\u2014 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\u2019s official provider uses fp4-quantized inference, which differs in performance from BF16 self-hosted deployments.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Tencent Hunyuan Hy4 vs. other comparable models<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>Model<\/th><th>Parameter count<\/th><th>Core advantages<\/th><th>Weaknesses<\/th><th>This Site\u2019s Rating<\/th><\/tr><\/thead>\n<tbody>\n<tr><td><strong>Hunyuan Hy4<\/strong><\/td><td>770B \/ 49B activated<\/td><td>1M context + image input + extremely low caching cost<\/td><td>Overly lengthy reasoning, delayed open-source weights<\/td><td>8.9<\/td><\/tr>\n<tr><td><a href=\"https:\/\/aidashxp.com\/en\/glm-5-3-review\/\">GLM 5.3<\/a><\/td><td>Unreleased<\/td><td>Robust million-token context inference, mature ecosystem<\/td><td>Average speed<\/td><td>8.4<\/td><\/tr>\n<tr><td><a href=\"https:\/\/aidashxp.com\/en\/kimi-k3-1-pro-review\/\">Kimi K3.1 Pro<\/a><\/td><td>2.8T<\/td><td>Largest open-source model, strong long-document reasoning<\/td><td>Pricing 5\u20138\u00d7 higher<\/td><td>8.8<\/td><\/tr>\n<tr><td><a href=\"https:\/\/aidashxp.com\/en\/mimo-v2-6-review\/\">Xiaomi MiMo V2.6<\/a><\/td><td>Unreleased<\/td><td>Three-tier pricing, Flash tier extremely affordable<\/td><td>Subpar performance on complex Agent tasks<\/td><td>8.1<\/td><\/tr>\n<\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">The conclusion is clear: Hy4 holds its ground as an \u201copen-source, 1M-context Agent model,\u201d priced competitively against GLM-series models rather than <a href=\"https:\/\/kimi.moonshot.cn\" target=\"_blank\" rel=\"nofollow noopener\">Kimi<\/a> the series. It is not the strongest, but<strong>Agent completion rate per unit cost<\/strong>It is currently the most competitive model in China\u2019s domestic AI ecosystem.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Overall Score<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>\u7ef4\u5ea6<\/th><th>Score<\/th><th>evaluate<\/th><\/tr><\/thead>\n<tbody>\n<tr><td>functional completeness<\/td><td>8.9 \/ 10<\/td><td>Supports 1M context length, image input, tool calling, and structured output.<\/td><\/tr>\n<tr><td>\u6613\u7528\u6027<\/td><td>8.6 \/ 10<\/td><td>OpenAI-compatible API, official vLLM\/SGLang Docker images\u2014integration requires virtually zero cost.<\/td><\/tr>\n<tr><td>Cost-effectiveness<\/td><td>9.3 \/ 10<\/td><td>Priced at $0.42\/$1.25 per million input\/output tokens; cache read costs only $0.021\u2014significant advantage in long-context scenarios.<\/td><\/tr>\n<tr><td>\u4e2d\u6587\u652f\u6301<\/td><td>9.2 \/ 10<\/td><td>Trained natively on Chinese corpora by a domestic team; delivers stable performance on Chinese tasks and terminology understanding.<\/td><\/tr>\n<tr><td>\u8f93\u51fa\u8d28\u91cf<\/td><td>8.7 \/ 10<\/td><td>Slightly outperforms GLM 5.3 and <a href=\"https:\/\/kimi.moonshot.cn\" target=\"_blank\" rel=\"nofollow noopener\">Kimi<\/a> K3 in blind evaluations\u2014but the margin is narrow.<\/td><\/tr>\n<\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Overall rating: 8.9\/10<\/strong><\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Frequently Asked Questions (FAQ)<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">Is Tencent Hunyuan Hy4 free?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Its API is not free\u2014it 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.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">What hardware configuration is required to deploy Hy4 locally?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">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 <code>vllm\/vllm-openai:hy4-preview<\/code> or the SGLang Docker image, <code>lmsysorg\/sglang:hy4-preview<\/code>enabled with MTP speculative decoding. It cannot run on consumer-grade GPUs in a single machine\u2014it is designed for data-center-level deployment.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Which should I choose: Hy4 or GLM 5.3?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">They are priced similarly and both support 1M context. Choose Hy4 for long-horizon programming agents or when image input is required; choose <a href=\"https:\/\/aidashxp.com\/en\/glm-5-3-review\/\">GLM 5.3<\/a>GLM 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\u2014the practical experience depends more on your prompts and toolchain.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Does Hy4 support Chinese\u2014and how well does it perform?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Yes\u2014Chinese 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.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Summary: Worth Adding to Your Model Pool<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Hy4 Stable Release Is Not \u201cYet Another Benchmark-Chasing Model\u201d\u2014It Solves a Very Real Problem:<strong>Delivering 1M-Context + Agent Tool Calling + Image Input at $0.40 per Million Tokens<\/strong>If your project is bottlenecked by long-context costs, it\u2019s worth running an A\/B test. But don\u2019t expect it to save you tuning effort\u2014its two inherent flaws\u2014overly verbose reasoning and excessive self-verification\u2014will genuinely erode part of its price advantage.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Want to Compare Real Pricing and Scores Across More Domestic and Overseas Models? Check Our <a href=\"https:\/\/aidashxp.com\/en\/ai-models\/\">AI Model Library<\/a> and <a href=\"https:\/\/aidashxp.com\/en\/compare-tools\/\">Tool Comparison Engine<\/a>or continue reading:<a href=\"https:\/\/aidashxp.com\/en\/glm-5-3-flashx-review\/\">GLM 5.3 FlashX benchmarking<\/a> \u00b7 <a href=\"https:\/\/aidashxp.com\/en\/kimi-k3-1-pro-review\/\">Kimi K3.1 Pro Review<\/a> \u00b7 <a href=\"https:\/\/aidashxp.com\/en\/particle-1-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<\/a> \u00b7 <a href=\"https:\/\/aidashxp.com\/en\/reviews\/\">All Reviews<\/a>.<\/p>","protected":false},"excerpt":{"rendered":"<p>8 \u6708\u5e95\u90a3\u7248\u5f00\u6e90\u6743\u91cd\u8fd8\u6302\u7740\u300cPrevie [&hellip;]<\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[6],"tags":[],"class_list":["post-695","post","type-post","status-publish","format-standard","hentry","category-ai-coding"],"_links":{"self":[{"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/posts\/695","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/comments?post=695"}],"version-history":[{"count":1,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/posts\/695\/revisions"}],"predecessor-version":[{"id":698,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/posts\/695\/revisions\/698"}],"wp:attachment":[{"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/media?parent=695"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/categories?post=695"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/tags?post=695"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}