A generated title card for 'Tencent Hy4 vs DeepSeek V4.1 Pro'. A left card for Tencent Hy4 reads 'Callable now — $0.834 / $2.501', '770B total / 49B active', 'Intelligence Index 35'; a right card for DeepSeek V4.1 Pro reads 'No endpoint', 'No model card', 'Named once in a release note'. A footer reads 'Index per Artificial Analysis v5.0.1; DeepSeek status checked 20 November 2026.'
Guides & Insights

Tencent Hy4 vs DeepSeek V4.1 Pro: Only One Has an Endpoint

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

DeepSeek V4.1 Pro has never been served to anyone. Not in preview, not in a trusted-access cohort, not behind a waitlist. The name appears exactly once in public, in DeepSeek's 10 September 2026 release note for DeepSeek V4.1 Flash, in a sentence about routing requests "until V4.1-Pro launches" — a plan DeepSeek withdrew after developers objected to their production traffic being swapped underneath them. Tencent Hy4, by contrast, has been generally available since 4 November 2026 and downloadable since 20 November, when Tencent published the weights under its own licence. So the useful comparison here is not the one the headline promises. It is Tencent Hy4 against the two DeepSeek models you can actually call, plus an account of what would have to appear before the promised third one entered the frame.

That framing is not a dodge. It is the only version of the comparison that can be priced, and the price is the interesting part: Tencent Hy4 at $0.834 per million input tokens and $2.501 per million output against DeepSeek V4 Pro 0813 at $1.32 and $3.96 on DeepSeek's own card. A model Tencent shipped sixteen days ago undercuts DeepSeek's current flagship tier on both directions while activating the same number of parameters per token.

Screenshot of DeepSeek's API pricing documentation listing two callable models, deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (V4-Pro-0813), with per-million peak and off-peak input, cache-hit and output rates, and no row for a V4.1 Pro tier.

The tier that has never shipped

Before anything else, the inventory. As of 20 November 2026, here is what exists for DeepSeek V4.1 Pro and what does not.

• No model card. DeepSeek's documentation index lists two callable models and two legacy aliases, all of which resolve to V4.1 Flash or V4 Pro. There is no page for a V4.1 Pro model.

• No price row. The rate card carries deepseek-flash at $0.30 per million input and $1.20 output at peak, and deepseek-v4-pro at $1.32 and $3.96. Nothing sits above V4 Pro.

• No weights. The deepseek-ai organisation on Hugging Face has published nothing after DeepSeek-V4.1-Flash. Given that DeepSeek has open-weighted every flagship it has shipped under MIT, a silent Pro tier would be a break with its own pattern.

• No independent score. Artificial Analysis returns a page for DeepSeek V4 Pro 0813 and for DeepSeek V4.1 Flash. The URL for a V4.1 Pro returns 404.

• One sentence of public commitment, since withdrawn. The September routing plan was the only place the name appeared, and the withdrawal removed its only product consequence.

None of that is an argument that DeepSeek is not building a Pro tier. The company describes V4.1 Flash as "the smallest model in our new architecture family," which only makes sense if larger members exist. It is an argument that you cannot put a name on a routing table.

Two models, 49 billion active parameters

Here is the coincidence that makes this pair worth comparing at all. Tencent Hy4 and DeepSeek V4 Pro are both sparse mixture-of-experts models that activate 49 billion parameters per token. They reach that number by very different routes, and the difference explains most of the behavioural gap.

• Total parameters — Tencent Hy4 is 770 billion across 78 layers, with 256 routed experts per layer and one shared expert, activating the top eight. DeepSeek V4 Pro is 1.6 trillion, more than twice the total for the same active count.

• Attention — Hy4 uses gated DeepSeek sparse attention with an IndexCache that reuses the indexer's selection across layers, plus a sliding-window path in the Flash variant. DeepSeek V4 Pro uses its own sparse attention implementation. The irony is that Tencent's card credits DeepSeek's architecture paper by name as an inspiration, so the two models are close relatives at the attention layer.

• Context and output — both serve 1M tokens of context. Hy4 returns up to 64K output tokens; DeepSeek V4 Pro returns up to 384K, six times more headroom in a single response.

• Modality — Hy4 reads text, images and video frames. DeepSeek V4 Pro is text-only; DeepSeek's own price sheet marks vision unsupported on the Pro row and supported on the Flash row.

• Draft model — Hy4 ships a 3.3B DFly2 speculative-decoding drafter that shares the target model's KV cache and sparse indices, which is how Tencent gets sustained decode above 150 tokens per second on its own infrastructure. DeepSeek publishes no equivalent component.

The consequence for a buyer is that the two models are not interchangeable servers behind the same interface. Hy4 will read a screenshot and answer with 64K on board. V4 Pro will read a 384K-output prompt batch and answer text-only. The overlap is text-in, text-out, long context — which is most work, and is where the price comparison actually bites.

What you can call today, priced

Three models are in play, and the rate cards differ in structure as well as level.

• Tencent Hy4 — $0.834 input, $2.501 output, $0.042 cached input. Flat pricing, no peak tier. Artificial Analysis records it at an Intelligence Index of 35, 93 tokens per second, and $1.03 per completed index task.

• DeepSeek V4 Pro 0813 — $1.32 input, $3.96 output at peak; $0.66 and $1.98 off-peak; cached input at $0.044 peak and $0.022 off. Artificial Analysis puts it at 23 on the index at maximum effort, 78 tokens per second, and describes it as below average in intelligence for its size.

• DeepSeek V4.1 Flash 1017 — $0.30 input, $1.20 output at peak; $0.15 and $0.60 off-peak; cached input at $0.006 peak. It is the cheap tier, activates 8 billion parameters during prefill and 16 billion during decode, and reads images.

The head-to-head that matters is the first two, and DeepSeek's peak/off-peak split decides it. At peak rates Tencent Hy4 undercuts V4 Pro in every direction: $0.834 against $1.32 on input, $2.501 against $3.96 on output, and $0.042 against $0.044 on cached reads. Off-peak the ordering reverses completely, because V4 Pro drops to $0.66 input, $1.98 output and $0.022 cached — 21% cheaper than Hy4 on input, 21% cheaper on output and half the cache rate. So the cheap model depends on when you run: a batch job that can be scheduled into DeepSeek's off-peak window is cheaper on DeepSeek, and one that runs when it runs is cheaper on Tencent.

The index gap runs the other way from what the parameter counts suggest. Hy4 scores 35, V4 Pro scores 23 at maximum effort. Twelve points is not a rounding difference; it is the difference between a model that can hold a multi-step agent task together and one that Artificial Analysis files below the median for open-weight models of its size. DeepSeek's own positioning of V4 Pro is a 1.6-trillion-parameter flagship, and the independent evaluation does not support treating it as one.

Where Hy4 wins, and where the price is a trap

The cheap model is not cheap on every axis, and pretending otherwise is how routing decisions go wrong.

• Win: cost per completed task. $1.03 for Hy4 against $1.77 for DeepSeek V4 Pro 0813 at maximum effort. Hy4 is 42% cheaper per completed index task, and that figure already absorbs its verbosity — 160 million output tokens across the index against DeepSeek's 119 million.

• Win: modality. Hy4's vision path is the difference between one model and two in a pipeline that has to read a PDF page and then reason over it. Adding a separate vision model to a DeepSeek stack costs more than the token difference.

• Win: licensing clarity. Both are open-weight — Hy4 under the Tencent Hy License 1.0 with its Section 10 model-as-a-service clause, V4 Pro under MIT — but Hy4 shipped its weights on 20 November with prebuilt vLLM and SGLang images, and DeepSeek's MIT licence is the more permissive of the two. For pure self-hosting freedom, DeepSeek wins that specific point.

• Trap: first-token latency. Hy4 takes 115.9 seconds to produce its first answer token on the index run, against a tier median of 12.7 seconds. DeepSeek V4 Pro is not fast either, but on any interactive surface the wait is the user experience, and neither model is built for it.

• Trap: output ceiling. 64K against 384K is not a small gap. If the artefact you want is a long generated file in one call, Hy4 cannot produce it and V4 Pro can.

• Trap: verbosity. Hy4 burns 160 million output tokens to complete the index, which is 2.5 times the tier median. On a per-token rate card that is fine; on a per-answer budget it erodes the price advantage by a third or more.

If V4.1 Pro ships, what changes

It is worth being concrete about the trigger, because "watch for a launch" is not a plan.

The first signal will be a model card in the deepseek-ai organisation, not a blog post. Every DeepSeek flagship has arrived weights-first or simultaneously, and the card carries the parameter count, the context length and the licence — the three facts a routing decision needs. A press release without a repository would be a weaker signal than the pattern implies.

The second is a rate-card row. DeepSeek prices on its own sheet before it prices anywhere else, and a Pro tier would sit above $1.32 input. That number is the one to compare against Hy4's $0.834, and it is hard to see how a 1.6-trillion-parameter-plus model lands below a 770-billion one on input pricing without DeepSeek deciding to buy share.

The third is an Artificial Analysis page. The V4.1 Pro URL is a 404 today. When it resolves, the comparison stops being architectural and becomes a table of two index scores at matched effort — which is the only form in which this matchup will ever be decidable.

Until all three exist, the honest description of DeepSeek V4.1 Pro is a name with a withdrawal attached. Pages that price it, benchmark it, or recommend it are describing a model that has no endpoint.

Running the comparison on one key

A generated two-column scoreboard titled 'Tencent Hy4 vs DeepSeek V4.1 Pro — what can be called'. Tencent Hy4 column: endpoint live, $0.834 / $2.501 per 1M tokens, 1M-token context, 64K max output, Intelligence Index 35, weights published 20 Nov 2026. DeepSeek V4.1 Pro column: no endpoint, no published price, context unstated, output limit unstated, no independent score, no weights. Footer: 'DeepSeek status verified against its own documentation on 20 November 2026; Tencent figures per Artificial Analysis v5.0.1.'

Tencent Hy4 and both callable DeepSeek models are on OrcaRouter under their own strings. Rather than argue about whether the Pro tier would beat Hy4, run the two that exist against your own tasks: tencent/hy4 at $0.834 and $2.501, deepseek/deepseek-v4-pro at $1.32 and $3.96, deepseek/deepseek-v4.1-flash at $0.15 and $0.60, each passed through at provider list price with zero markup.

The routing DSL does the part a benchmark cannot. A weighted route that sends most traffic to the cheaper model and escalates to the other on a failed check is how you discover whether the twelve-point index gap is worth 37% more per token for your specific workload, and it costs one afternoon rather than a migration.

When—or if—a Pro tier appears, it is a model-string change on the same key rather than a second contract. That is the practical argument for holding the seat open: the cost of being ready is a line in a config file.

What the evidence supports

Screenshot of the Hugging Face model card for tencent/Hy4 showing the repository header, download and like counts, the MoE and multimodal tags, the Tencent Hy License badge and the links to ModelScope, GitCode and GitHub.

Strip the headline away and there are two defensible claims here, plus one that cannot be made.

Tencent Hy4 is callable, open-weighted since 20 November, cheaper per token than DeepSeek's flagship tier at peak rates in every direction, cheaper per completed index task by 42%, and twelve points ahead on the independent index. That is a real result for a model released sixteen days earlier, and it is the reason the open-weight release matters more than the launch did.

DeepSeek V4.1 Pro is a tier name with no specification attached. Nothing about it can be priced, benchmarked or recommended, and the only public statement about it has already been walked back. The comparison a buyer can run today is Hy4 against V4 Pro and V4.1 Flash, and on that comparison Hy4 wins on score and on cost per completed task, V4 Pro wins on output headroom, and DeepSeek wins on price whenever the work can be scheduled off-peak.

The claim that cannot be made is that Hy4 beats the Pro tier. Nobody has measured it, because there is nothing to measure. The useful stance is to treat the name as a placeholder and revisit this page when DeepSeek puts a card, a rate and an index page behind it.

Compared in this article4

Detected from this article · Benchmarks: Artificial Analysis · updated daily