Tencent Hy4 preview review (2026): an honest look at the open-source 770B coding and reasoning model — strengths, limits, pricing, and creator fit.
Hy4 preview is a serious open-source release: a 770B mixture-of-experts model with a 1M-token context, tuned hard for coding, agentic engineering, and long-context analysis, with open weights and cheap API access. For developers and research-heavy work it looks strong on paper — though the headline benchmarks are Tencent's own. For creators, understand what it is: a text-only drafting and reasoning brain, not a content or publishing tool, and one tuned for engineering rather than marketing voice.
Most of the reviews on this site cover content and publishing tools. Hy4 preview is neither, and it would be misleading to score it as if it were. It is a large language model — Tencent Hunyuan's next-generation MoE model, released and open-sourced on August 28, 2026 — built for software engineering, office and financial analysis, game development, and scientific research. So this review judges it honestly as a text model a creator or creator-developer might reach for, then draws the line where its job ends and a content engine's begins.
The technical shape is real and impressive: 770 billion total parameters with about 49 billion active per step, a context window exceeding 1 million tokens, open weights (plus an FP8 variant) on Hugging Face, ModelScope, GitCode, and CNB, and API pricing that undercuts most Western frontier models. Tencent also reports an unusual agentic result — the model participating in optimizing its own training and inference — which is notable for the field even if it isn't a feature you use directly.
The honest caveats are about evidence and scope. The strongest performance numbers, including a 2.99/4.00 internal blind evaluation, are Tencent's own and await independent verification. And Hy4 is tuned for engineering, not English marketing copy, so treating it as a caption or hook writer is a category mismatch. Below, the scores reflect it as a coding-and-reasoning LLM, with a clear note on where creators need a different tool entirely.
Hy4 preview is a text-in, text-out large language model from Tencent's Hunyuan team, released and open-sourced on August 28, 2026. Architecturally it is a mixture of experts — 770 billion total parameters with roughly 49 billion active per token — with a context window Tencent describes as exceeding 1 million tokens. It targets productivity and technical work: debugging and long-context coding, office and financial analysis, game development, and scientific research such as AI R&D and physics simulation. Tencent published open weights and a lower-precision FP8 variant on Hugging Face, ModelScope, GitCode, and CNB. What it is not is a content creation platform. Hy4 produces text and code — no images, no video, no audio, no captions, no scheduling, no publishing. It exposes strong reasoning and long-context abilities that make it a capable front end for drafting and analysis, but every output is raw text on a screen. Turning that text into finished, on-brand, published content is a separate job it does not attempt.
Hy4 preview fits developers, research-heavy teams, and technically comfortable creators who want a cheap, openly available reasoning model for coding, agentic tasks, and digesting large documents. Its million-token context is genuinely useful for anyone who needs to reason over an entire transcript, codebase, or corpus in one pass, and the open weights matter for teams with data-governance or self-hosting requirements. It is a weak fit for creators who assumed a "next-generation model" would produce finished marketing content — because it is engineering-tuned, its short-form voice is unproven, and it outputs raw text, so all of the captioning, formatting, brand-styling, and publishing work is still ahead of you.
| Dimension | Score | Why |
|---|---|---|
| Coding & agentic capability | 4.0 / 5 | Purpose-built for debugging, long-horizon coding, and agentic tasks, with strong reported benchmarks — though the numbers are Tencent's own. |
| Long-context reasoning | 4.3 / 5 | A 1M-token window lets it ingest and reason over whole transcripts, codebases, and document sets in one pass. |
| Open-source access & self-hosting | 4.3 / 5 | Open weights plus an FP8 variant on Hugging Face, ModelScope, GitCode, and CNB — real portability and privacy control. |
| API pricing & value | 4.0 / 5 | Roughly $0.834/M input and $2.501/M output undercuts most Western frontier models; a limited free window at launch. |
| Creative & marketing writing | 2.8 / 5 | Tuned for engineering and analysis, not English short-form voice; caption/hook quality is unproven and reads like raw model output. |
| Western availability & ecosystem | 3.5 / 5 | Accessible via OpenRouter and open weights, but first-party products are Tencent-centric and some teams weigh data-governance factors. |
| Benchmark transparency | 3.0 / 5 | Headline scores come from Tencent's internal blind evaluation; independent third-party results were not yet available at review time. |
| Content & publishing capability | 1.0 / 5 | None by design — no images, video, captions, multi-format output, brand voice, or publishing. It stops at text. |
On pure model economics, Hy4 preview is aggressively priced. Tencent lists API access at roughly $0.834 per million input tokens and $2.501 per million output tokens through Tencent Cloud TokenHub and OpenRouter, with a limited free-access window via WorkBuddy and CodeBuddy at launch. For a model at this scale with a million-token context, that input price in particular is low relative to most Western frontier options — attractive for high-volume drafting or large-document analysis. Confirm current numbers on Tencent Cloud, since launch pricing and free windows change.
The open weights add a second pricing story: if you have the hardware, you can run Hy4 (or the FP8 variant) yourself and pay in compute rather than per token, which matters for privacy-sensitive or high-throughput workloads. The catch is that a 770B MoE is not a laptop model — meaningful self-hosting means real GPU capacity, so the "free" weights carry an infrastructure cost most individual creators won't want to absorb.
For a content workflow, the honest framing is that Hy4's price is the cost of raw text, which is only one line item. Whatever you draft 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 $99/mo), which is a different unit than per-token text, and comparing the two directly is a category error.
| Use case | Fit | Why |
|---|---|---|
| Debugging and long-horizon coding | Strong | This is the model's core target, with strong reported coding and agentic benchmarks. |
| Reasoning over a huge transcript or codebase | Strong | The 1M-token context ingests very large inputs in one pass for analysis or summarization. |
| Cheap, high-volume text drafting via API | Strong | Low per-token pricing makes it economical for bulk drafting and research synthesis. |
| Self-hosting for privacy or governance | OK | Open weights enable it, but 770B parameters demand serious hardware even in FP8. |
| Writing on-brand captions and hooks | Weak | It is engineering-tuned, so short-form marketing voice is unproven and reads like raw model output. |
| Making captioned video, carousels, or images | Weak | Text-only by design — it generates no visual media and no feed-ready assets. |
| Publishing content across platforms | Weak | There is no scheduler or publisher; it produces text and stops. |
Kompozy is not a competitor to Hy4 preview, and pretending otherwise would be dishonest — Kompozy is not a large language model and does not publish open weights or reason over a codebase. They sit at different points in the workflow. Hy4 answers "how do I draft, code, or analyze 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 draft. Feed it the text Hy4 produced (or just a 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 cheap, open, long-context 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 engineering, and long-context analysis, it is a compelling open-source option — a 770B MoE model with a 1M-token context, open weights, and low API pricing. Just note the strongest benchmarks are Tencent's own. For creating finished marketing content it is the wrong category: it outputs raw text only and is tuned for engineering, not short-form voice.
Both are large, open-leaning coding and reasoning models, and Tencent benchmarks Hy4 against DeepSeek V4 Pro directly, reporting an edge on some engineering tests. Those numbers are vendor-reported, so treat head-to-head claims cautiously until independent evaluations land. In practice, pick based on your own task tests, licensing terms, and hosting needs rather than a single benchmark.
No. Hy4 is a text-in, text-out model — it produces text and code only, with no images, video, audio, captions, or publishing. To turn its output into visual, multi-format, scheduled content you pair it with a generation-and-publishing engine like Kompozy.
Tencent released the Hy4 preview model weights openly, along with a lower-precision FP8 variant, on Hugging Face, ModelScope, GitCode, and CNB. Confirm the specific license and usage terms on Tencent's release pages before deploying it commercially.
API access was listed at roughly $0.834 per million input tokens and $2.501 per million output tokens via Tencent Cloud TokenHub and OpenRouter, with a limited free window through WorkBuddy and CodeBuddy at launch. Self-hosting the open weights trades per-token cost for substantial GPU hardware. Verify current figures on Tencent Cloud.
It can draft and reason over large 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 drafting and research 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.
See Tencent Hy4 preview vs Kompozy comparison → · Get Started →