A title card headlined 'Qwen4-Max vs Tencent HY4 Preview: one you can download, one you can only rent', with three chips reading HY4 Preview under Apache 2.0 at 770B total and 49B active, Qwen4-Max closed weights with a 2.4T backbone and 71B active, and a price per million input of $0.834 against $1.50, captioned to the vendor model cards, Vals AI and Artificial Analysis on 17 November 2026.
Guides & Insights

Qwen4-Max vs Tencent HY4 Preview: One You Can Download, One You Can Only Rent, and Neither Number Means What It Looks Like

Author

Elias Hawthorne

Date Published

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

T​encent HY4 Preview has been downloadable since 28 August 2026 under Apache 2.0 — 770 billion parameters in total, 49 billion activated per token, 78 layers, 256 routed experts plus one shared, a native multi-token-prediction layer for speculative decoding, and a 1,048,576-token context window. Qwen4-Max has been rentable since 16 November 2026 — 2.4 trillion backbone parameters, 71 billion activated per token, plus 328 billion of N-gram embedding parameters, and no checkpoint of any kind. Both are Chinese flagship sparse mixture-of-experts models with million-token windows. One you can hold, one you can only call, and at $0.834 against $1.50 per million input tokens the one you can hold is also the cheaper of the two on the vendor's own card. The interesting part is why the price gap exists at all, and what each side's published numbers leave out.

The two cards, side by side

A two-column scoreboard comparing Qwen4-Max and Tencent HY4 Preview: proprietary licence against Apache 2.0; 2.4T with 71B active against 770B with 49B active; $1.50 in and $4.50 out against $0.834 and $2.501; 1,000,000 context against 1,048,576; 131,072 maximum output against 64,000; text, image and video against text only. Footer records HY4 Preview at 53.21% on the Vals AI index, 23rd of 33, and Qwen4-Max at 43 on the Artificial Analysis index, 38th of 81.

• Licence — Apache 2.0 with published weights for HY4 Preview vs proprietary, no weights, no download for Qwen4-Max

• Size — 770B total / 49B active for HY4 Preview vs 2.4T backbone / 71B active plus 328B of N-gram embeddings for Qwen4-Max

• Price per million tokens — $0.834 in / $2.501 out vs $1.50 in / $4.50 out

• Cache read — $0.042 per million vs $0.075

• Context window — 1,048,576 tokens vs 1,000,000, of which 983,616 usable as input in thinking mode

• Max output — 64,000 tokens vs 131,072 non-thinking, or 262,144 shared with reasoning

• Input modalities — text only vs text, image and video

• Reasoning levels — high, low and none, defaulting to high vs low, high and xhigh, defaulting to high

• Independent index score — 53.21% on the Vals AI index, 23rd of 33 vs 43 on the Artificial Analysis Intelligence Index, 38th of 81

Read the modality row before the price rows. HY4 Preview is a text-in, text-out model. Qwen4-Max reads video. Those two facts do not sit on the same axis, so for any team whose workload includes frames the comparison ends at that row and the rest of this is academic.

A screenshot of the Hugging Face model card for tencent/Hy4-preview, showing the repository header, the Apache 2.0 licence tag, the safetensors and mixture-of-experts tags and the start of the evaluation results section.

Apache 2.0 is not the same offer at 770 billion parameters

Weights you can download are weights you can inspect, fine-tune, quantise and serve on your own terms, and that is a real difference from a model behind someone else's API. It is also, at this scale, a difference available to very few teams. A 770-billion-parameter model with 49 billion active will not run on anything short of a multi-node accelerator cluster, and the published FP8 checkpoint that accompanied the preview exists mainly so that inference providers and research groups with fabric access can serve it. Apache 2.0 removes the legal barrier, not the hardware one.

Which is why the licence's practical value shows up somewhere other than self-hosting. It shows up as competitive supply: because anyone can serve HY4 Preview, multiple platforms do, and price discipline follows. Qwen4-Max has no such mechanism. There is one deployment, run by A​libaba, and every platform quoting a Qwen4-Max price is quoting A​libaba's price with whatever margin it can add. That is why its Global-region card and Singapore card differ by 5.8% and nothing else differs by much — there is no third-party capacity to undercut it.

The number to distrust on each card

T​encent's launch material includes an internal blind evaluation — 163 experts, 203 engineering tasks — in which HY4 Preview averaged 2.99 against GLM 5.3 at 2.92 and Kimi K3 at 2.94. That is a vendor-run study comparing three models, published by the vendor that finished first, on a scoring scale it defined. It is not evidence of much beyond the direction of the team's own testing, and T​encent was careful to publish it alongside a list of known rough edges: the model spends longer than necessary reasoning and over-verifies its own work.

On the independent side, Vals AI places HY4 Preview at 53.21% on its index, 23rd of 33 models tested. On Vals' own Terminal-Bench 4.0 run, the preview scores 8.08%. T​encent's materials quote terminal-agent numbers in the eighties — on a different revision of the benchmark, under a different harness, which is why the two figures are not a contradiction so much as a demonstration that agentic benchmark scores are a property of the harness as much as the model. Anyone quoting either number without the revision attached is doing the reader a disservice.

Qwen4-Max has the opposite problem. Its vendor figures are unusually detailed — Work360 at 52.69 against Qwen3.8-Max's 47.74, average cost per task $4.43 down to $1.67, secure-code function-level pass rates 22.0% to 31.8% — and none of them name a competitor. What A​libaba published is a self-comparison against its own previous generation, which is the most defensible kind of vendor benchmark and the least useful for choosing between vendors. The only cross-vendor reading is Artificial Analysis, which scores Qwen4-Max at 43 and records no speed measurement for it.

Where the 45% price gap goes

HY4 Preview costs $0.834 in and $2.501 out; Qwen4-Max costs $1.50 and $4.50. That is 45% cheaper on input and 44% cheaper on output — near enough to a flat discount, which is what you would expect from two models priced against similar hardware economics. The gap is not free: HY4 Preview's 64,000-token output ceiling is half of Qwen4-Max's non-thinking figure, its context is text-only, and the two models deliver different reasoning control. Qwen4-Max also gives you three effort levels against T​encent's high, low and none, though both default to their top setting, and both fold unfamiliar effort strings rather than rejecting them.

The other place the discount can evaporate is verification behaviour. T​encent documents that the preview over-verifies; Qwen4-Max's independent token total is 127 million output tokens across the Intelligence Index against a median of 78.6 million, which Artificial Analysis calls very verbose. Both models, in other words, are chatty, and a per-token discount on a model that spends more tokens than its competitor recovers less than the rate card promises. Measure output volume per task before you model the saving.

One model is a preview, the other is a general release

This is the asymmetry that gets lost whenever the two names appear in the same sentence. HY4 Preview shipped in August and T​encent has since moved the line forward — the Hy4 general-availability release in November added native vision and computer use, and there is a separate Hy4 Flash in the family. The preview string is what is available with open weights, and it is a preview because T​encent said so, with the rough edges documented.

Qwen4-Max is a general release, on a metered card, in six regions, with a preview generation that is being retired on 25 November. Those are different maturity stages, and comparing them on benchmark scores flatters whichever one the harness happened to suit.

A screenshot of the OrcaRouter model page for tencent/hy4-preview, showing the 1M-token context badge, the 64K maximum output, the model identifier and the start of the pricing panel.

Running both without picking a winner first

OrcaRouter routes T​encent HY4 Preview at $0.834 in and $2.501 out per million with cached input at $0.042 — T​encent's own card, passed through at 0% markup — through the same key and the same OpenAI-compatible endpoint as Claude Opus 5.5, GPT-6.1 Sol, Grok 4.5 and Qwen3.8-Max. Qwen4-Max is not in the catalogue; the Qwen4 generation has not been added yet, so that side of the comparison is a direct A​libaba credential.

Which is the arrangement this pair actually rewards. Put the open-weight model behind the routing layer where failover applies, keep the closed one on a direct key while you establish its behaviour, and decide on your own traffic: send text-only work to the route that costs 45% less and streams from several providers, and send anything with images or video to the model that can read them. The routing DSL makes that a rule on modality rather than a decision someone has to remember, and model fusion — a panel of models answering together — is the cheap way to see whether the two disagree on the jobs you care about before you commit either of them to production.

What the weights argument is really about

A​libaba's launch post ends by saying smaller Qwen4 models are coming to open source. That sentence is the shape of the next year: the open tier of Chinese flagship families will keep landing a generation behind the closed one, Apache-licensed and downloadable, and teams will keep choosing between the model they can hold and the model that is six months ahead.

For now the choice is unusually clean. HY4 Preview is the honest open option at this scale, priced accordingly, with a text-only surface and vendor numbers that need a harness attached to mean anything. Qwen4-Max is the newer closed option, better specified on every axis a sheet can measure, with one independent score that is lower than its launch framing implies and no latency number at all. Neither card is the whole story, and the 45% between them is smaller than the difference in what each one can be pointed at.