Unlocking Tencent Hy4: The Quiet Chinese Flagship That Outscores Inkling 2 Small — But the Weights Aren't Here Yet

On November 4, 2026, three model launches sat on the Hacker News front page at once: Moonshot's Kimi K3.1 Pro weights first, Ricursive's Ripple-65B second, Tencent's Hy4 third. The crowd had already decided what mattered: downloadable weights.

That was the wrong ranking. A day later, the first independent score landed: Tencent Hy4 scored 35 on the Artificial Analysis Intelligence Index — above Inkling 2 Small, level with Grok 4.7, and the second-best score of any Chinese lab, behind only Kimi K3.1 Pro.

But here is the twist: you still cannot download Hy4. The weights are promised, not shipped. So is this the open-weight flagship Tencent claims, or an announcement dressed as a release? Here is what is real, what it costs, and what is still missing.

What Tencent actually shipped on November 4

Hy4 is the full release of Tencent's flagship, successor to the text-only Hy4 preview of August 28, 2026, which was open from day one. The full release adds two things the preview never had: vision, and computer use — the ability to operate a graphical interface and a terminal.

It went live the same day through Tencent Cloud's TokenHub and OpenRouter (Tencent Cloud is currently the only provider), and it powers Tencent's own apps, WorkBuddy and CodeBuddy, which offered a free quota from November 4 to 15 — demand was so strong Tencent expanded the Hy4 inference cluster tenfold on November 5.

The launch post gave no parameter count. The preview was a 770-billion-parameter mixture-of-experts model with 49 billion active per token, but Tencent has published no size, licence or model card for the final release — so don't let anyone tell you the final Hy4 is 770B. That number belongs to the preview.

The score that matters: 35 on the Artificial Analysis Intelligence Index

The number to remember is 35 — Hy4's score on the Artificial Analysis Intelligence Index v5, published November 5, a composite of nine evaluations covering reasoning, knowledge, coding, planning and tool use.

That puts it second among Chinese labs — behind only Kimi K3.1 Pro at 46, ahead of Qwen3.8 Max and DeepSeek V4.1 Flash, level with Grok 4.7 at its highest setting, one point behind Claude Haiku 5.5 at 36, and four clear of Inkling 2 Small's 31.

The sub-scores are more useful than the headline: 31% on FrontierCode 2.0 (agentic coding), 39.39% on Terminal-Bench 4.0 (15th of 60 models), and 33% on HLE-Diamond, the hardest slice of Humanity's Last Exam. Artificial Analysis prices the whole index run at about $0.78 per task on Hy4 — roughly a tenth of the same run on Grok 4.7.

The honest catch: the weights aren't out yet

Now the part Tencent would rather you skim past: the full Hy4 weights are not downloadable. Tencent says it "plans to open-source" them, with a system card to follow — the press release promises both "in the coming days." No date was given.

As of launch, the evidence was blunt: the Hugging Face page for the final model returned an error, the GitHub repository did not exist, and the only downloadable Hy4 checkpoint was the August preview under Apache 2.0. None of that is a download link.

This is the same story as Inkling 2 Small: a flagship announced as open-weight, with the weights still behind a promise. Two of November's most talked-about "open" models are API-only right now — the difference is that Tencent shipped a downloadable preview in August. Thinking Machines has shipped no weights at all.

A model is open-weight when you can download it, not when a press release says you will be able to. By that rule, Hy4 is a hosted model with an open-weight preview attached — for now.

What it costs — and the discount that halves it

Pricing is refreshingly clean — unchanged from the preview: $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens.

Then the discount: on Tencent Cloud TokenHub, Hy4 is 50% off through December 31, 2026 — $0.417 in, $1.2505 out per million tokens, as on OpenRouter.

Two things to plan around. First, the output-to-input ratio is about 3:1 and Hy4 is a heavy reasoner, so your bill leans toward the $2.501 output meter, not the headline $0.834. Second, Kimi K3.1 Pro charges $3.00 in and $15.00 out at list price — Hy4 is roughly a quarter of Kimi's input price and a sixth of its output price, for a model trailing Kimi by 11 points on the same index. Serious value, and the same pricing trap the GPT-6 Cyber story warned about applies here: the headline rate is never the bill.

Hy4 vs Inkling 2 Small vs Ripple-65B vs Kimi K3.1 Pro

	Tencent Hy4	Inkling 2 Small	Ripple-65B Preview	Kimi K3.1 Pro
Artificial Analysis Index	35	31	Not yet scored	46
Weights today	Promised, not out	Promised, not out	Apache 2.0, downloadable now	Apache 2.0, downloadable now
Price per 1M in / out	$0.834 / $2.501 ($0.417 / $1.251 discounted)	$0.30 / $1.00 (Tinker only)	$0.20 / $0.60	$3.00 / $15.00
The catch	Weights promised, size undisclosed	Weights promised, Tinker-only	Smallest, newest lab	Best score, by far the priciest

The pecking order is clear: Kimi is the best model and the most expensive, Ripple is the one you can actually download today, and Hy4 is the value play — the second-best independent score among Chinese models at a price that embarrasses everything near it. Inkling is the odd one out: weaker than Hy4 on the independent index, still locked inside Tinker.

What Hy4 is actually good at

Tencent's chart shows Hy4 leading on long-context and refactoring tests like SWE Atlas Refactoring and OneMillionBench, and its Terminal-Bench 4.0 score puts it in the top quarter of the 60 models tested — the agentic coding workload most teams actually run.

Computer use is the other differentiator: the first model in its class to pair vision with GUI and terminal control, so it can drive the software your team already uses instead of only answering questions about it — and with a 1-million-token context window (per OpenRouter), it can hold an entire codebase or a week of documents in one session.

Where it falls short

The ugliest number is the hallucination rate: on AA-Omniscience v1.1, Hy4 managed 50% accuracy while hallucinating 89% of the time. Where a wrong answer is worse than no answer, that number should set your verification budget.

Then the over-thinking: Tencent warns Hy4 spends longer than necessary reasoning and over-verifies its own work. Artificial Analysis quantified it — 160 million output tokens across its index run, against a 75 million median, more than double a typical model. That means slower answers and a bigger bill on the $2.501-per-million output meter.

And the ceiling is real: on Tencent's own chart, Claude Opus 5.5 wins all 12 tests, and Hy4's 33% on HLE-Diamond trails the closed frontier by a wide margin. Tencent acknowledges a "large gap" to the top closed models. This is a value flagship, not a frontier flagship — and there is nothing wrong with that, as long as nobody sells it as the other thing.

A 7-step playbook: try Hy4 this week without wasting money

To evaluate Hy4 properly, treat it like a trial, not a migration.

1. Start on TokenHub while the 50% discount lasts (through Dec 31, 2026).
2. Grab the free WorkBuddy or CodeBuddy quota before November 15.
3. Pick one job it is built for — a large refactor, a long-document review, a multi-step office workflow.
4. Cap the reasoning with a max output or reasoning budget per call — the over-thinking is manageable if you fence it in.
5. Verify everything it claims — with an 89% hallucination rate, treat every factual claim as a draft until checked.
6. Cache aggressively — at $0.042 per million cached input tokens, repeated prompts over the same documents are nearly free.
7. Decide with numbers, not excitement: cost per finished task, error rate, time saved — and don't build infrastructure around weights that don't exist yet.

Final takeaway: the quiet flagship, with an asterisk

Tencent Hy4 is the most interesting model of November that nobody is shouting about: it outscores the Western open-weight hope on the independent index, ties Grok 4.7 at a tenth of the cost, adds vision and computer use, and undercuts Kimi K3.1 Pro by a factor of four on price. Where cost per task matters most, it is the best value on the board right now.

But the asterisk is the whole story: weights promised, not published; parameter count undisclosed; hallucination rate ugly. And the model that beats it by 11 points — Kimi — is downloadable today.

My honest take: use Hy4 through the API this week, while the discount and free quota last. It will probably earn a place in your agent stack on price alone. Just don't call it open-weight until you can download it, and don't build anything that depends on the weights arriving on schedule. Tencent earned the benefit of the doubt with its August preview — now it must close the gap between the press release and the download page.