Tencent Hy4 Review 2026: The Open AI Model That Can See and Use Your Computer
Tencent Hy4 review 2026: the official November 4 release adds visual reasoning and computer use to a long-context mixture-of-experts model — with the Preview-versus-official distinction explained, self-run benchmarks flagged as such, very low API pricing, and what's still pending.
Tencent Hy4 review 2026: before anything else, one point of precision that saves a lot of confusion — "Hy4" currently refers to two different releases. Hy4 Preview, announced August 28, 2026, shipped with open weights. The official Hy4 launched November 4, 2026, adding visual understanding and computer use, but Tencent said its full weights and system card would follow in the days after the announcement. Writing about the official release as if its weights are already downloadable, or assuming it shares Preview's license, would be an error this review is built to avoid.
What makes the official release worth covering regardless is the direction it signals. Preview was a long-context, text-focused model aimed at coding and productivity. The official version adds image and chart understanding, and the ability to operate command-line and graphical interfaces — a move from a model that answers to a model that can see, analyze, and act. Tencent frames it as a step toward long-running, multi-step agent work: planning, executing, and verifying results across a task rather than ending after a single reply.
Almost every performance claim in the launch materials comes from Tencent itself or from evaluations run with Tencent-affiliated teams. That doesn't make the claims wrong. It means this review labels them as claims throughout, and treats the strongest-sounding numbers as starting points for your own testing, not settled facts.
Preview vs. Official — Don't Confuse Them
| Item | Hy4 Preview | Hy4 Official |
|---|---|---|
| Announced | August 28, 2026 | November 4, 2026 |
| Weights | Published (Hugging Face) | Promised "in the following days" — not confirmed at check date |
| Context | 1M tokens | ~1.05M tokens in Tencent's service |
| Parameters | 770B total (MoE) | No final specs found confirming the architecture is unchanged |
| Vision | Not equivalent to official | Image and chart understanding added |
| Computer use | Not the focus of the first announcement | New announced capability |
| License | Apache 2.0 | Can't be confirmed before weights are published |
The practical rule: anything about weights, local deployment, or the Apache 2.0 license applies to Preview unless and until Tencent publishes the official weights and a license for them. This review keeps that line sharp in every section below.
What's New in the Official Release
Image & chart understanding
The official version analyzes images, handles fine visual detail, reads charts and diagrams, and tackles vision-dependent math problems — a meaningful addition over Preview's text-first focus.
Computer use
Tencent says Hy4 can work with command-line and graphical interfaces — navigating between programs and executing steps — though the announcement doesn't prove it controls arbitrary software without setup, tooling, and human oversight.
Long-horizon tasks
Improvements to planning, execution, and verification across tasks that don't end in one response — relevant to coding, data analysis, research, and workflow automation.
Coding & game development
Specialized training on tools like Blender, CAD, Unreal Engine, Unity, and Godot, with Tencent reporting 19.1% gains on a GameDev test and 9.8% on GameLogic versus Preview.
Scientific research tasks
Work across math, physics, environmental science, and engineering, evaluated in part through TenureTrail — 123 executable tasks across seven disciplines.
Those architecture numbers describe Preview, per its published model card, not a confirmed spec sheet for the official release. The design is a mixture-of-experts: 78 layers, the first dense and the other 77 MoE, with 8 routed experts plus the shared expert activated per token, Gated DeepSeek Sparse Attention with IndexCache reusing attention indices across layers, and an extra MTP layer to speed decoding. The point of the design is that a model with hundreds of billions of parameters doesn't activate them all on every step — but that doesn't make it light to run. Memory and multi-accelerator distribution remain serious challenges.
Benchmarks — Whose Tests Are These?
The strongest number in the launch is a financial-tasks evaluation Tencent ran with Tencent Financial Technology: 579 realistic finance tasks, judged with 213 finance professionals and 153 external experts, roughly 70% of whom had more than five years of experience. On a 4.00 scale, Tencent reports Hy4 at 3.04, DeepSeek V4.1 Flash 1017 at 2.97, and both Kimi K3.1 Pro and GLM-5.3 at 2.68. That places Hy4 first among the open models in that evaluation, per Tencent — but it's not a general ranking of all models, and the methodology isn't published in enough detail to reproduce independently.
Tencent also reports its internal hallucination rate falling from 12.95% in Preview to 7.2% in the official release, and errors from unsupported assumptions dropping from 23.1% to 5.6%. Those are Tencent's internal evaluation figures — useful as a before-and-after on Tencent's own test, not a statement that Hy4 hallucinates 7.2% of the time in your use case.
A separate result comes from the Preview announcement: in a blind evaluation by 163 expert reviewers across 203 engineering tasks (WorkBuddy), Preview scored 2.99 out of 4, against 2.94 for Kimi K3 and 2.92 for GLM-5.3. That belongs to Preview, not the official release, and shouldn't be mixed into one table with the official model's numbers without labeling the version.
API Pricing & Launch Offers
| Usage type | Price per million tokens |
|---|---|
| Input | ~$0.834 |
| Output | ~$2.501 |
| Cached input | ~$0.042 |
Tencent says the official Hy4 keeps Preview's pricing; on Tencent Cloud's Chinese pricing the equivalent is 6 yuan input, 18 yuan output, and 0.3 yuan cached input per million tokens. As a worked example, a coding task with 100,000 input tokens and 20,000 output tokens, with no caching, would cost roughly $0.133 at those rates — a theoretical figure before taxes or fees, and likely an underestimate for agent workflows, where multiple steps, repeated context, and tool calls consume more tokens.
Launch offers are time-limited, and "free" applies to them, not to the API: a 50% discount on Hy4 through Tencent Cloud TokenHub until December 31, 2026; WorkBuddy and CodeBuddy free from November 4 to 15, 2026; and a cheaper Hy4 Flash variant announced as coming soon but not available at launch. The API itself is pay-as-you-go.
How to Access It — and Whether You Can Run It Locally
For hosted use, Tencent Cloud TokenHub exposes the model as hy4 through interfaces compatible with OpenAI Chat Completions, Responses, and Anthropic Messages, per Tencent's documentation. WorkBuddy handles office and multi-step file tasks, CodeBuddy targets developers, and Tencent lists OpenRouter among access channels — though whether the official Hy4 appears in its model list should be checked before linking to it directly. A minimal call follows the OpenAI-compatible pattern:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_TENCENT_TOKENHUB_KEY",
base_url="https://api.hunyuan.cloud.tencent.com/v1"
)
response = client.chat.completions.create(
model="hy4",
messages=[{"role": "user",
"content": "Analyze this codebase and identify its main architectural risks."}]
)
print(response.choices[0].message.content)That's illustrative, based on the base URL in Tencent's documentation, and requires a TokenHub account and valid key — never commit a real key to a public repository.
Cybersecurity & Safety Claims
Reuters reports that Tencent focused its cybersecurity training on defensive uses — finding software vulnerabilities, analyzing security issues, helping fix weaknesses — while excluding training material tied to building intrusion tools, breaking into systems, or probing systems to exploit them. Tencent also says its internal expert testing showed no increase in the model's ability to carry out cyberattacks or bypass protections, and found no meaningful uplift for harmful biological experiments.
Those are the company's conclusions, not independent attestation that the risks are absent. Still, the framing matters: a large model heading toward open-weight availability carries a real responsibility to assess misuse risk before the weights are out in the world, and Tencent putting that assessment in the launch narrative is a point in its favor. How that compares with other labs' safety postures for comparably capable open models is worth tracking as independent evaluators weigh in.
Pros & Cons
Hy4 vs. Kimi K3.1 Pro vs. DeepSeek
~1.05M context, 770B total / 49B active (Preview specs), vision and computer use in the official release, API at ~$0.834 / $2.501 per million tokens. Preview weights under Apache 2.0; official weights pending.
1M context, 2.8T total / 104B active per its model card, published weights under Kimi's own license, and a standard API price previously verified at about $2 input / $10 output — a figure to re-check on Kimi's pricing page before publishing any final price comparison.
Scored 2.97 against Hy4's 3.04 in Tencent's own financial-tasks evaluation — a narrow, self-run gap that doesn't establish a general winner between the two.
There's no clear overall winner here. Hy4 is the most attractive on API cost; Kimi's advantage is official weights that could actually be downloaded as of this review's date. The right choice turns on real task performance, license constraints, deployment cost, and the specific project — none of which a launch-week comparison can settle.
Who Should Use It
Ideal user: a developer or team that wants to test a large, low-cost, long-context model through an OpenAI-compatible API, or a research or enterprise team ready to work with Preview's Apache 2.0 weights while waiting for the official release's weights and system card.
Look elsewhere if: you need confirmed, downloadable official weights today, require independently verified benchmarks before committing, or want to run a frontier-scale model on modest hardware. And test Arabic, English, and other languages separately — an international API doesn't prove equal performance across languages.
Expert Editorial Opinion
The most useful framing for Hy4 isn't "another big open model." It's the combination Tencent is attempting: a very long context window, visual understanding, and the ability to operate computers, in a model priced low enough that using it for long, multi-step agent work doesn't immediately become a budgeting problem. Each piece alone is increasingly common; the bundle, at this price, is what makes the launch interesting.
The Preview-versus-official split is the part of this story most likely to trip up coverage, and it's worth restating plainly. The weights, the Apache 2.0 license, and the architecture details are all Preview facts. The vision and computer-use capabilities are official-release claims. A single sentence that blends the two — "Hy4 is open-weight under Apache 2.0 and can use your computer" — is wrong in a way that would mislead a developer planning around it, and until Tencent publishes the official weights and license, that sentence shouldn't be written.
On the evidence, the financial-tasks result deserves a fair reading. Hundreds of real tasks judged by hundreds of professionals is a more serious evaluation design than many launch benchmarks, and a 3.04 versus 2.68 gap over Kimi and GLM is not trivial. It's also a test run with Tencent's own financial-technology arm, with methodology not fully open to reproduction. The honest reading: a strong, specific, credible-looking result that earns a place on a shortlist, not a ruling on which model is best at finance, let alone everything else.
The safety framing is the part that most distinguishes this launch from a pure capability announcement. Training the cybersecurity capability toward defense, excluding offensive-tool material, and publicly describing misuse testing before weights go out is the right instinct for a model of this scale. It remains Tencent's self-assessment, and it will matter more once independent evaluators have the official weights in hand — which, at the time of writing, they don't yet.
Put simply: this is a promising release from a company that appears to be taking both capability and responsibility seriously, with the most important verification steps — official weights, a system card, independent benchmarks — still ahead of it. That's reason enough to watch closely, test cheaply through the discounted API, and hold off on any architecture decision that depends on the unpublished parts.
Final Verdict
Tencent Hy4 earns a solid score for a coherent capability set — long context, visual understanding, computer use — at an unusually low API price, with Apache 2.0 Preview weights already available and a safety narrative that goes beyond the usual. The score is held back by how new and unverified the official release is: weights, system card, and license still pending, nearly all benchmarks self-run or affiliated, and computer use announced rather than independently shown. For API-based experimentation it's an easy, cheap thing to try now; for anything depending on official open weights, it's too early to commit.
| Dimension | Weight | Score /10 | Why |
|---|---|---|---|
| Technical quality | 30% | 8.0 | A strong, coherent architecture and capability set with credible-looking results, though evidence is largely self-run or Tencent-affiliated |
| Price-to-value | 25% | 8.5 | Very low API pricing, cached-input discounts, and a 50% launch discount make heavy long-context use unusually affordable |
| Maturity & documentation | 20% | 5.5 | The official release is a day old, with weights and system card pending, though Preview is documented in detail |
| Ceiling & flexibility | 15% | 7.5 | Apache 2.0 Preview weights, OpenAI/Anthropic-compatible APIs, and vLLM/SGLang support, offset by enormous hardware demands and an unconfirmed official license |
| Honesty of positioning | 10% | 7.0 | Tencent clearly separates Preview from official and states what's pending, but headline benchmark and safety conclusions are its own |
Weighted total: (8.0×0.30) + (8.5×0.25) + (5.5×0.20) + (7.5×0.15) + (7.0×0.10) = 7.45/10, rounded to 7.5/10 — Score band: 7.0–7.9, "Competent but compromised — promising and unusually affordable, with the key verification steps still to come."
❓ Frequently Asked Questions
🔗 Related ToolRadar Reviews
Official sources: Hy4 launch release · Hy4 Preview on Hugging Face · TokenHub API guide · Preview announcement. Check whether the official weights, system card, and license have been published before making any deployment decision.
Comments
Post a Comment