> ## Documentation Index > Fetch the complete documentation index at: https://docs.apiyi.com/llms.txt > Use this file to discover all available pages before exploring further. # Tencent Hunyuan Hy4 Is Live on APIYI > Tencent's Hunyuan Hy4 is now available on APIYI: a 770B MoE model with a 960K input window and the first Hunyuan release with vision and Computer Use. Priced at $0.85 / $2.55 per 1M tokens in Default and SVIP. ## Key Highlights * **Official release is live**: `hy4` is available on APIYI through the OpenAI-compatible `/v1/chat/completions` endpoint in the `Default` and `SVIP` groups * **770B MoE, 49B active**: each inference activates roughly 6.4% of the parameters, decoupling knowledge capacity from per-call cost * **960K input / 64K output**: the official context claim exceeds one million tokens, while the TokenHub limit is 960K input and 64K output * **First Hunyuan model that sees and acts**: vision understanding and Computer Use arrive together, covering graphical interfaces and command-line software operation * **Hallucination rate nearly halved**: down from 12.95% to 7.2%, with guess-induced errors falling from 23.1% to 5.6% ## Background Since rebuilding its infrastructure in February 2026, Tencent Hunyuan has shipped a major version roughly every two months: a Preview first, then a formal release shaped by user feedback. Hy4 preview launched on August 28, 2026 as a model "optimized for agent, coding, and productivity scenarios." On November 4, the formal Hy4 release arrived, scaling from the sub-300B Hy3 series to 770B total parameters and, for the first time, putting vision and Computer Use into a single model. This is not a benchmark-only launch. The six areas Tencent highlights — office collaboration, long-running personal assistance, financial analysis, front-end development, game development, and scientific research — each come with internal evaluations or real business data. What matters more for developers is the change in call shape: mixed text-and-image tasks no longer require chaining a separate multimodal model, because text generation, deep reasoning, and multimodal understanding live in one model and the chain collapses from multiple steps to one. APIYI now serves the formal release under the model ID `hy4`. Projects that already have a text chat pipeline can swap the `model` field and run an evaluation pass. ## Deep Dive ### Specs: 770B MoE with a 960K context | Spec | Hy4 (official release) | | - | - | | Architecture | Mixture of Experts (MoE) | | Total parameters | 770B | | Active parameters | 49B | | Max input | 960K tokens | | Max output | 64K tokens | | Official context claim | Over 1 million tokens | | Optimization focus | Agent, coding, productivity | Active parameters account for roughly 6.4% of the total, a direct result of the MoE design: inference only wakes part of the expert network, so knowledge capacity and per-call cost are decoupled. In practice, you get a large model's knowledge breadth without paying a price proportional to its total size. A 960K input window means a mid-sized code repository or a hundred-page contract can go in at once, with no chunking layer of your own. The 64K output ceiling defines how much long-form code, structured reporting, or complete documentation the model can produce in one pass — in real development work, that output cap is usually hit before raw parameter count matters. ### Two firsts: reading images and taking action > [card] Vision understanding > > Hy4 reads images for the first time, covering complex chart interpretation, document understanding, multi-image analysis, and math-heavy visual reasoning. Mixed text-and-image tasks used to require a separate multimodal model; now one model and one call handle them. > [card] Computer Use > > Vision applied to action: operating software through graphical interfaces and the command line. Hunyuan trained specifically on professional tools such as Blender and CAD and on long-running cross-application tasks, validating them repeatedly in a sandbox. Vision input is confirmed working on APIYI: pass images directly in `messages`, with no separate multimodal model. Computer Use is a new upstream capability — validate it with a single request before relying on it. Tencent shared two near-limit vision cases: locating a roughly two-meter rooftop component inside a 2.8-billion-pixel aerial photo, and reassembling a 900-piece jigsaw puzzle into Monet's *Bordighera*. The first tests fine-grained localization at extreme resolution; the second tests holistic understanding and spatial reasoning across visual fragments. Computer Use comes with infrastructure: Hunyuan and the Tencent Cloud Agent Runtime team built a sandbox that deploys hundreds of applications for the model to execute against repeatedly. That sandbox is also open to enterprises and developers for training and running their own agents. ### Measured results across six scenarios | Scenario | Official figure | | - | - | | Office collaboration (WorkBuddy) | 83.18 average score, 96.15% task pass rate, first place in China on 9 of 12 scenarios | | Financial crowdsourcing test | 213 experts and 579 real tasks, 3.04 points — first among open models | | Game development | GameDev up 19.1% and GameLogic up 9.8% versus preview | | Scientific research | First among open models on TenureTrail (7 disciplines, 123 tasks) | The financial comparison set includes DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68), and GLM-5.3 (2.68). Its value lies in the "real tasks plus outside experts" design: 153 of the 213 experts came from leading external financial institutions, 70% with more than five years of experience, covering fund research, securities analysis, quantitative strategy, asset allocation, and accounting. Two research cases deserve a closer look. Facing 84 unordered two-dimensional slices containing 6.99 million cells — including mirrored flips — Hy4 completed slice alignment and 3D reconstruction in 1.5 hours, offline, with no specialized registration tool and no cell-type labels, and beat the default pipeline of common registration tools without human intervention. Separately, in work with a Peking University mathematics team, Hy4 found a counterexample to the Méndez–Pinto–Villarreal conjecture, proposed a new condition to refine it, and verified the related proof in Lean — a full loop from counterexample discovery to proposition revision to proof verification. ### Reliability: hallucination and guessing rates > [!NOTE] > > Different Tencent channels report slightly different figures (another disclosure cites 15.76% down to 5.45%), but all agree that guess-induced errors fell from 23.1% to 5.6%. The figures below follow the official self-built test disclosure. * Hallucination rate: 12.95% → 7.2% * Guess-induced error rate: 23.1% → 5.6% For finance, legal, and retrieval-augmented applications, these numbers may matter more than any leaderboard: the model leans toward answering from evidence instead of completing a plausible-sounding fact. ## Practical Use ### Code example The example below calls `hy4` through the APIYI gateway with the OpenAI Python SDK. Set `APIYI_API_KEY` in your environment first: ```python theme={null} import os from openai import OpenAI client = OpenAI( api_key=os.environ["APIYI_API_KEY"], base_url="https://api.apiyi.com/v1" ) response = client.chat.completions.create( model="hy4", messages=[ {"role": "system", "content": "You are a senior financial analyst. Lead with the conclusion, then list the evidence."}, {"role": "user", "content": "Turn this quarterly report into a table of revenue, gross margin, and operating cash flow, then name the one point most worth pressing on."} ], stream=True ) for chunk in response: if chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end="") ``` Both streaming and non-streaming calls return in the standard OpenAI format. With a cache hit, repeated prefixes bill at \$0.045 per 1M tokens, so keep fixed system prompts and long documents at the front of the request. ### Where it fits > [card] Long documents and code > > The 960K input window suits contract review, technical documentation, and mid-sized repositories read in a single pass, without chunking and reassembly. > [card] Agents and tool calling > > Optimized for long-horizon tasks: multi-step tool calls, cross-application flows, and automation agents that need sustained execution. > [card] Finance and research > > First among open models on the financial test; combined with the lower hallucination rate, a fit for research summaries, valuation models, and due-diligence Q\&A. > [card] Front-end and game development > > Deep support for UE, Unity, and Godot, with double-digit percentage gains in GameDev and GameLogic versus preview. ### Usage notes 1. **Vision is available now; validate Computer Use first**: pass images directly in `messages` to use vision understanding, and image tokens bill as input. Computer Use is a new upstream capability, so verify it with a single request in your own format before depending on it in production. 2. **Put stable content first**: cache hits cost about 5.3% of the input price, so keep system prompts, long documents, and conversation history at the front. 3. **Sandbox long-running tasks**: run Computer Use workflows in an isolated environment before granting production access. 4. **Budget by `usage`**: long contexts and image input raise real token consumption, so do not estimate cost from the unit price alone. ## Pricing and Availability ### Pricing | Billing item | Price per 1M tokens | | - | - | | Input (shown as "prompt price" in the console) | \$0.85 | | Output (shown as "completion price" in the console) | \$2.55 | | Cached input (read) | \$0.045 | > [!NOTE] > > Pricing is aligned with the official list and may change; the table is for reference only. The **Model Pricing** tab in the top navigation is authoritative: [Model Pricing](https://docs.apiyi.com/en/models/index.md). Tencent's official Hy4 pricing is 6 CNY input, 18 CNY output, and 0.3 CNY cached input per million tokens. At APIYI's fixed 1:7 rate, the two are effectively the same. Tencent Cloud TokenHub also runs a limited-time list-price discount, which is an upstream direct-purchase offer independent of APIYI's pricing. ### Available groups `Default` and `SVIP` are both open. The model ID `hy4` is identical in both and uses the OpenAI-compatible `/v1/chat/completions` endpoint. `Default` works for initial integration checks; workloads that need higher stability and concurrency can follow the existing group strategy and select `SVIP`. ### Stacking recharge promotions The prices above are list rates and can be combined with recharge bonuses and group discounts. Actual charges follow the console. See [recharge promotions](https://docs.apiyi.com/en/faq/recharge-promotions.md) for the rules. ## Summary and Recommendations The headline for Hy4's formal release is not that parameters grew from under 300B to 770B. It is that three things happened at once: the model can see images, it can operate a computer, and its hallucination rate was nearly halved. That moves it from a text productivity model to an agent model that can touch real work objects. * **Long documents and code understanding**: 960K input with 64K output — a candidate to replace chunk-based pipelines first * **Finance, research, and due diligence**: first among open models on the financial test, and more trustworthy with the lower hallucination rate * **Agents and GUI automation**: validate Computer Use stability in a sandbox, then widen permissions gradually * **Cost-sensitive, high-volume short text**: Hy4 costs more than Hy3, so Hy3 remains the better choice for that workload > [!NOTE] > > **Sources and verification date**: > > * Tencent Hunyuan Hy4 formal release notes: `cloud.tencent.cn/developer/article/2776227` > * Hy4 capability upgrades and billing: `cloud.tencent.com.cn/developer/article/2780277` > * Tencent Cloud Hunyuan product page: `cloud.tencent.com/product/tclm` > * APIYI live model registry: `docs.apiyi.com/model-registry.json` > * Research and pricing verified on November 10, 2026 This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.