Tencent Hy4 vs Tencent HY3: Four Times the Context, Four and a Half Times the Price

Tencent HY3 is a 295-billion-parameter mixture-of-experts model you can download, serve on your own hardware and call for $0.18 per million input tokens. Tencent Hy4 is a 770-billion-parameter model with four times the context window, a desktop-driving computer-use mode, and an input price of $0.834 — with its weights promised rather than published. They are one generation and four months apart in the same Hunyuan line, and the useful comparison is not which scores higher on Tencent's own sheet. It is which one you are allowed to run, where, and at what finished cost.

That framing matters because the two models are not competing for the same slot. HY3 is a workhorse you can put behind a firewall, fine-tune and forget about. Hy4 is a frontier-class API model that shipped this morning, whose licence has not been written yet. Choosing between them is mostly a question about constraints, and only secondarily about capability.

The two models, dimension by dimension

• Total parameters — 295B for Tencent HY3 vs 770B for Tencent Hy4.

• Active parameters per token — 21B vs 49B, both mixture-of-experts with top-8 routing. HY3 has 192 experts across 80 layers; Hy4 has 256 routed experts plus one shared across 78 layers, with a 10B multi-token-prediction layer for speculative decoding.

• Context window — 262,144 tokens vs 1,048,576. Exactly four times, and the single biggest functional difference between them.

• Input modality — text only on HY3 vs text and image on the GA build, which is the first time the line has had vision at all.

• Weights — Apache 2.0 and downloadable today for HY3, including an FP8 checkpoint. For Hy4, planned, undated, unlicensed.

• Price on OrcaRouter — $0.18 input and $0.59 output per million tokens for HY3; $0.834 and $2.501 for Hy4, which is 4.6 times on input and 4.2 times on output.

• Cached input — $0.059 per million for HY3 vs $0.042 for Hy4. The expensive model is the cheaper one here, and that single line changes the arithmetic for agent workloads more than the headline ratio suggests.

The cache line is where the price gap halves

Most comparisons stop at the input price and conclude that Hy4 costs four and a half times as much. For a chat box that is roughly right. For an agent that re-reads the same 200,000-token repository on every turn, it is wrong by a factor of two, because the two models bill cached tokens in the opposite order to their fresh ones.

Take a twenty-turn session over a 200,000-token context, 2,000 output tokens per turn, with 90% of input served from cache. That is 3.6 million cached input tokens, 400,000 fresh input tokens and 40,000 output tokens.

• Tencent HY3 — $0.21 for cached input, $0.07 for fresh input, $0.02 for output. Total $0.31.

• Tencent Hy4 — $0.15 for cached input, $0.33 for fresh input, $0.10 for output. Total $0.58.

The same session with no cache at all comes to $0.74 on HY3 and $3.44 on Hy4 — the full 4.6 times. Caching is what closes the gap, and it closes it from both ends: Hy4's cache rate is 29% below HY3's, while its fresh rate is 4.6 times above. On a long-running agent the bill lands at roughly 1.9 times, not 4.6.

Two caveats before anyone budgets on that. The cache-hit price only applies to tokens the provider actually serves from cache, and hit rates depend on how you structure the prefix; a session that edits its own context every turn will cache less than one that appends. And a 200,000-token context fits inside HY3's window, which is why the comparison is possible at all — past 262,144 tokens HY3 is not a cheaper option, it is not an option.

What the vendor numbers say, and the two-run problem

Tencent published a comparison table when it launched the Hy4 preview, and it is the only side-by-side the company has ever put out. Read it as a direction of travel, not as a set of deltas.

• Terminal-Bench 2.1 — 70.8 for HY3 against 85.4 for the Hy4 lineage, in Tencent's own comparison chart.

• DeepSWE — 28.0 against 64.3, the largest relative jump in the table.

• SWE-bench Pro — HY3 at 57.9; the Hy4 lineage at 65.7.

• BrowseComp — 84.2 for HY3, not restated for Hy4.

• Humanity's Last Exam with tools — 53.2 for HY3, not restated.

The trap is the 70.8. Tencent's HY3 launch post reported Terminal-Bench 2.1 at 71.7, and the Hy4 comparison chart prints 70.8 for the same model name. That is almost certainly two runs of the same benchmark with the harness or sampling changed, and it means you cannot subtract the two published rows and claim a 14.6-point improvement. The honest version is that the Hy4 lineage scores in the mid-eighties on a benchmark where HY3 scores around 71, from two separate vendor runs that have never been reconciled in public.

The GA release did not republish any of this. It added two internal comparisons instead — GameDev +19.1% and GameLogic +9.8% over the preview — and a finance expert panel where Hy4 averaged 3.04 against 2.68 for both Kimi K3.1 Pro and GLM-5.3. None of it is independent, and none of it compares directly against HY3. If you need a number to defend a procurement decision, neither model has one that survives scrutiny; what exists is four months of practitioner reports on HY3 and one morning of vendor claims on Hy4.

Where HY3 still wins outright

Four things, and the first three are not close.

It runs where you run. Apache 2.0 weights, an FP8 checkpoint, vLLM and SGLang recipes, and a 295B total size that fits on a single multi-GPU node at FP8. Hy4 needs a cluster most teams do not have, and cannot legally be self-hosted at all until Tencent publishes a licence.

It is faster per token and per task. A 21B-active model with a 256K window does less work per call than a 49B-active model that reasons longer and, in the preview, over-verified its own output. Tencent claims the GA build shortened the chain of thought, but even a fixed Hy4 is doing more computation per answer than HY3. For latency-sensitive paths — autocomplete, chat, classification — HY3 is the right shape.

It is cheap enough to be the default. At $0.18 and $0.59, HY3 is in the band where you route everything to it and escalate selectively. At $0.834 and $2.501, Hy4 is a model you route to deliberately.

It has a track record. HY3 has been generally available since 6 July 2026, with three months of production use and a preview period before that. Hy4 GA is hours old and shipped without a benchmark table. A model with a known failure profile is worth more than a better model with an unknown one.

Where Hy4 is the only answer

Three workloads genuinely do not have a HY3 option.

• Anything over 262,144 tokens. A 400-page contract, a monorepo, an hour of transcripts — HY3 cannot hold it, and chunking is a different product with different failure modes.

• Images. HY3 is text in, text out. Hy4 GA reads charts, diagrams and screenshots, which makes it the first Hy model that can sit in a document or UI pipeline at all.

• Desktop and application control. Computer use through a combined GUI and CLI is new in the GA build, and no amount of prompting gives HY3 a mouse.

Everything else is a cost decision, and the cost decision is closer than the rate card implies once caching is in play.

Running both without choosing

The sensible architecture for most teams is not a migration. It is a split: HY3 as the default for short, frequent, latency-sensitive work, and Hy4 for the long-context, visual or agentic slice that HY3 physically cannot serve.

On OrcaRouter that split is a routing rule rather than two integrations. Both models are on one key — orca passes the provider's list price through with no markup, so a Tencent price change is live here the day it happens rather than repriced by us — and tencent/hy3 and tencent/hy4-preview are both callable today. Moving a workload between them is a model-string change, not a second contract or a new SDK, and a routing DSL can send anything above a token threshold or carrying an image to the larger model while everything else falls through to the cheaper one. For the GA build specifically, the catalogue does not list it yet; when it lands it will sit behind the same key, and until then the preview is the closest routable proxy at the identical price.

Automatic failover is the other half of that argument. A brand-new flagship with no independent benchmarks and a launch-day queue is exactly the case where you want a second model behind the route, and HY3 — self-hostable, published, three months in production — is the obvious fallback for anything that is not long-context or visual.

The decision, stated plainly

If your work fits in 256K tokens, arrives as text, and runs at volume, Tencent HY3 is still the better buy on 4 November 2026. It is four to five times cheaper on fresh tokens, roughly twice as cheap on a realistic cached agent session, faster, downloadable, and known.

If you need a million-token window, image understanding or desktop control, Tencent Hy4 is not competing with HY3 — it is the only Hunyuan model that does the job, and the comparison becomes Hy4 against whatever else you would use for that workload. Its price is high in the family and unremarkable against the frontier, and its real handicap today is not cost. It is that the model has no public benchmark table, no licence, and one morning of existence.