Tencent Hy4 review (2026): an honest verdict on the full 770B model with native vision and Computer Use — strengths, limits, pricing, and creator fit.
The full Hy4 is a meaningful step up from the preview: the same 770B mixture-of-experts base, now with native visual understanding and Computer Use — it can read a chart or screen and act on software through graphical and command-line interfaces. For engineering, analysis, and agentic work it looks strong, though the headline numbers are Tencent's own and the open weights were still pending at launch. For creators, the category is unchanged: it is a reasoning-and-vision brain that outputs text and actions, not a tool that produces or publishes finished content.
Most reviews on this site cover content and publishing tools. Hy4 is neither, and scoring it as one would mislead. It is Tencent Hunyuan's full next-generation large language model, released November 4, 2026, following the Hy4 preview that was open-sourced on August 28. So this review judges it honestly as a multimodal reasoning model a creator or creator-developer might reach for, then draws the line where its job ends and a content engine's begins.
What's new over the preview is the reason to look again. On top of the 770B MoE base, the full Hy4 introduces native visual understanding — reading image detail, analyzing complex charts, handling visual math — and Computer Use, executing tasks through graphical and command-line interfaces with training that reaches tools like Blender and CAD. That makes it a model that can both see a screen and operate one, which is a genuinely different shape than a text-only LLM.
The honest caveats are about evidence and scope. The strongest performance figures — an 83.18 WorkBuddy average, a lower financial hallucination rate than the preview — are Tencent's internal results and await independent verification, and Tencent's own comparisons are mainly against other open-source models, not closed frontier systems. On openness, Tencent positioned Hy4 as an open-source model but said weights and a System Card would follow "soon," so at launch it was API- and product-access only, unlike the preview whose weights are already public. The scores below reflect it as a multimodal reasoning model, with a clear note on where creators need a different tool entirely.
Hy4 is a multimodal large language model from Tencent's Hunyuan team, released November 4, 2026. It keeps the mixture-of-experts architecture of the Hy4 preview, which the series puts at 770 billion total parameters, and adds two capabilities new to the line: native visual understanding (reasoning over images, charts, and screens) and Computer Use (operating software through graphical and command-line interfaces, with specialized training on tools like Blender and CAD that Tencent says improves its ability to plan and carry out long, complex tasks). It is aimed at software engineering, office and financial analysis, game development, and research — productivity and technical work, not chat or creative media. What it is not is a content creation platform. Even with vision, Hy4's outputs are text and actions — no images, video, audio, captions, scheduling, or publishing. The visual ability is for understanding inputs, not generating media. Its strength is digesting large, technical, or now visual material and reasoning over it, which makes it a capable front end; turning that into finished, on-brand, published content is a separate job it does not attempt.
The full Hy4 fits developers, research-heavy teams, and technically comfortable creators who want a capable reasoning model that can also read visual inputs and drive software — code debugging, document and chart analysis, and agentic, multi-step tasks. The vision upgrade is useful for anyone who works from diagrams, dashboards, or screenshots rather than clean text, and Computer Use appeals to builders automating tool-driven workflows. It is a weak fit for creators who assumed a "vision model" would make visual content — because it understands images rather than generating them, outputs only text and actions, and is tuned for engineering, so all of the captioning, formatting, brand-styling, and publishing work is still ahead of you. Teams that specifically need open weights should also note those were not yet released at launch.
| Dimension | Score | Why |
|---|---|---|
| Vision & multimodal understanding | 3.8 / 5 | New to the line: reads image detail, analyzes complex charts, and handles visual math — strong on paper, but the results are vendor-reported. |
| Computer Use & agentic execution | 3.9 / 5 | Operates graphical and command-line interfaces, with specialized Blender/CAD training Tencent says aids long, complex tasks; notable, though independent evaluation was not yet available. |
| Coding & engineering | 4.0 / 5 | Carries the preview's core strength in debugging and long-horizon coding, with Tencent reporting gains on game-dev tasks. |
| Long-context reasoning | 4.0 / 5 | Tencent says long-context comprehension improved over the preview, though it did not restate an exact window at the full launch. |
| Open access & self-hosting | 3.5 / 5 | Positioned as open source, but weights plus a System Card were promised "soon" and not released at launch — API and Tencent-product access only for now. |
| API pricing & value | 4.0 / 5 | Reported unchanged from the preview at ~$0.83/M input and ~$2.50/M output, with a 50% TokenHub promo through year-end. |
| Benchmark transparency | 3.0 / 5 | Headline scores are Tencent's internal evaluations; independent third-party results for the vision and Computer-Use claims were not yet available. |
| Content & publishing capability | 1.0 / 5 | None by design — no media generation, multi-format output, brand voice, or publishing. It understands and reasons; it does not make or ship content. |
On model economics, Hy4 stays aggressive. Tencent says API access is unchanged from the preview — about $0.90 per million input tokens and $2.70 per million output tokens, with cached input at about $0.045 per million tokens — through Tencent Cloud TokenHub and OpenRouter, and Tencent Cloud TokenHub was billing at 50% through December 31, 2026. For a frontier-scale multimodal model, that input price in particular is low relative to most Western options, which is attractive for high-volume drafting, document analysis, or now chart and screenshot reading. Confirm current numbers on Tencent Cloud, since promos and launch windows change.
The open-weights story that made the preview compelling is, for now, a promise rather than a download: Tencent said Hy4's weights and a System Card would follow soon, so at launch there is no self-hosted, pay-in-compute option for the full model the way there is for the preview. If data governance or self-hosting is your reason for choosing Hy4, that gap matters — either wait for the weights or use the already-open preview.
For a content workflow, the honest framing is that Hy4's price is the cost of raw reasoning, which is one line item. Whatever it drafts or analyzes still needs captioning, formatting into video and carousels, brand-voice governance, and publishing — and those steps, not the tokens, are where a content operation's real cost and time live. A content engine like Kompozy is priced by generated-and-published output (credit-based tiers from $199/mo), a different unit than per-token reasoning, so comparing the two directly is a category error.
| Use case | Fit | Why |
|---|---|---|
| Reasoning over charts, dashboards, or screenshots | Strong | Native visual understanding lets it read and analyze visual inputs, not just text. |
| Debugging and long-horizon coding | Strong | It carries the preview's core engineering strength, with reported game-dev gains. |
| Agentic, tool-driven automation | OK | Computer Use operates software via GUI and command line, though independent reliability data is not yet available. |
| Self-hosting for privacy or governance | Weak | The full-model weights were not released at launch; use the open Hy4 preview if self-hosting is the requirement. |
| Writing on-brand captions and hooks | Weak | Engineering-tuned, so short-form marketing voice is unproven and reads like raw model output. |
| Making captioned video, carousels, or images | Weak | It understands images but generates none — no visual media or feed-ready assets. |
| Publishing content across platforms | Weak | There is no scheduler or publisher; it produces text and actions and stops. |
Kompozy is not a competitor to Hy4, and pretending otherwise would be dishonest — Kompozy is not a large language model, does not reason over a codebase or a chart, and does not operate your computer. They sit at different points in the workflow. Hy4 answers "how do I understand, draft, or automate this?" Kompozy answers "how do I turn this into published, on-brand content across platforms?" In fact Kompozy's own copy generation runs on Claude and OpenAI, with the Persona Brief and banned-word filters shaping voice, so the model layer is deliberately abstracted away from the creator.
Where Kompozy earns its place is everything after the analysis. Feed it the brief Hy4 produced from a visual or technical source, and it generates across 18 formats — Persona and HeyGen avatar video that narrates the idea on camera, Clipped Shorts with branded captions, Carousels via HyperFrames, Photo Posts, Quote Graphics, Blog Articles, Email Newsletters, and Text Posts — each held to one brand voice, then reframed per platform and scheduled and published across eight social platforms plus blog and email behind a per-post review. If your need is a capable multimodal reasoning model, Hy4 is a legitimate pick; if it is finished content shipped everywhere in your voice, that's Kompozy's lane, and it's a different tool for a different job.
For coding, agentic automation, and now reasoning over visual inputs like charts and screens, the full Hy4 is a compelling option — a 770B MoE model with native vision and Computer Use at low API pricing. Note the strongest benchmarks are Tencent's own and the open weights were pending at launch. For creating finished marketing content it is the wrong category: it outputs text and actions only and is tuned for engineering, not short-form voice.
The full Hy4 (November 4, 2026) keeps the preview's 770B mixture-of-experts base and adds two capabilities new to the line: native visual understanding — reading image detail, charts, and visual math — and Computer Use, operating software through graphical and command-line interfaces. The August preview is text-focused and already has public open weights; the full model's weights were promised but not yet released at launch.
No. Hy4's vision is for understanding inputs, not generating them — it reads and reasons over images and charts but outputs text and actions, with no image, video, or audio generation. To turn its analysis into visual, multi-format, scheduled content you pair it with a generation-and-publishing engine like Kompozy.
Tencent positioned Hy4 as an open-source model and said the weights, with a risk-focused System Card, would be released soon — but at launch access was via API (Tencent Cloud TokenHub, OpenRouter) and Tencent products, not a public download. If you need open weights immediately, the Hy4 preview's are already published. Confirm the full model's weights timeline and license on Tencent's pages.
Tencent says API pricing is unchanged from the preview: about $0.90 per million input tokens and $2.70 per million output tokens, with cached input at about $0.045 per million tokens, via Tencent Cloud TokenHub and OpenRouter. Tencent Cloud TokenHub was billing at 50% through December 31, 2026. Verify current figures on Tencent Cloud.
It can draft and reason over large or visual inputs well, but it is optimized for engineering and analysis, so its English short-form and marketing voice is unproven and its raw output reads like model output. For creators it works best as an upstream reasoning and briefing brain, with a content engine handling voice, formatting, and distribution.
Kompozy takes a draft or source and generates 18 formats — persona/avatar video, carousels, images, quote graphics, blogs, and newsletters — in one brand voice, then schedules and publishes across eight social platforms plus blog and email with autopilot and a per-post review.
Hy4's vision is new and vendor-benchmarked, while Claude and GPT models have more mature, independently tested multimodal and creative-writing abilities. Pick based on your own task tests, licensing and hosting needs, and whether voice quality or low per-token cost matters more, rather than a single launch benchmark.