Tencent’s Hy4 Did Not Climb the Chart — It Arrived Halfway Up It

Tencent’s Hy4 did not climb OpenRouter’s weekly chart. It arrived at No. 10, carrying 4.01 trillion tokens in its first measured week while the preview version remained at No. 6 with another 6.58 trillion. That is not proof that Hy4 is the tenth-best language model in the world. It is something more useful for engineering teams: evidence that model traffic can migrate at production speed, long before independent evaluations have settled whether the new endpoint is actually better.

The two Hy4 entries together account for 10.59 trillion tokens in OpenRouter’s trailing seven-day window. If they were one SKU, that volume would sit behind only DeepSeek’s two leading variants and GLM 5.3 Flash. Combining them is analytically imperfect—the preview and production release have different capabilities, and some traffic may represent deliberate A/B testing—but the split reveals the shape of adoption. Tencent did not have to persuade every developer to start a new application. It gave existing users a nearby model identifier and a credible reason to try it.

That distinction matters because OpenRouter counts tokens, not users, applications, revenue, task success or satisfaction. Four trillion tokens could represent a broad base of small products or a handful of high-volume workloads. It cannot tell us whether Hy4 writes better code, completes more tool calls, or costs less per successful task. It can tell us that developers were willing to route real workloads through a model whose final weights and system card have not yet been released. Distribution has compressed the model-adoption cycle from quarters to days.

The production label is a migration event, not a coronation

Tencent’s launch case leans heavily on internal evaluations. The company reports an 83.18 average on WorkBuddy and a 96.15% pass rate, placing first in nine of twelve mainland-China scenarios. In a blind finance evaluation covering 213 professionals and 579 real-world tasks, Tencent says Hy4 scored 3.04, ahead of DeepSeek V4.1 Flash 1017 at 2.97 and Kimi K3.1 Pro and GLM 5.3 at 2.68. It also reports gains of 19.1% on GameDev and 9.8% on GameLogic over the preview.

Those are useful hypotheses, not independent verdicts. Vendor-organized benchmarks routinely encode choices about prompts, graders, retries and failure handling that may not resemble a team’s production environment. The right response is neither to dismiss the results nor to repeat them as settled fact. Treat them as a map of where Tencent expects Hy4 to be strongest: office workflows, finance, game development, visual understanding, and GUI or terminal computer use. Then test those claims against your own traces.

The most consequential number in the launch material is not a leaderboard score. Tencent says hallucinations fell from 12.95% with Hy4 Preview to 7.2%, while errors caused by unsupported assumptions dropped from 23.1% to 5.6%. For an agent, unsupported assumptions are often more expensive than a plainly wrong answer. A model that invents application state, silently chooses a permission boundary, or claims a tool succeeded can create a clean-looking failure that survives downstream checks.

But Tencent has not published a reproducible harness, confidence intervals or enough detail to know how those categories were assigned. Teams should build a compact acceptance suite from failures they already own: stale-data questions, unavailable tools, ambiguous permissions, conflicting instructions, and cases where the correct response is to stop and ask. Score abstention, recovery and tool-state fidelity separately from final-answer quality. A model that produces a prettier answer after an imaginary tool call is not an upgrade.

A 770-billion-parameter model can be open and still not be local

The preview model card describes a 770-billion-parameter mixture-of-experts architecture with 49 billion parameters active per token, 78 layers, 256 routed experts plus one shared expert, top-eight expert activation, and a one-million-token context window. The Apache 2.0 preview weights can be served with vLLM or SGLang, which makes the release genuinely inspectable and modifiable. Yet “open weights” and “practical self-hosting” remain two different checkboxes.

Tencent’s official FP8 deployment recipe uses tensor parallelism across eight GPUs. Sparse activation reduces compute per token, but it does not make the full expert set disappear from memory, remove networking costs, or eliminate the operational work of batching, routing and capacity planning. For most product teams, an OpenRouter or Tencent Cloud endpoint is the realistic evaluation path. Self-hosting becomes rational when data residency, sustained utilization, model modification, or predictable capacity can pay for the infrastructure and on-call burden.

The final Hy4 weights and system card are still promised rather than delivered. That should be a governance gate for teams that require reproducible artifacts, documented safety behavior, or internal model approval. Tencent says the model was trained on defensive security material while offensive exploit content was excluded, and its security labs found no significant offensive-capability concern in evaluated areas. That is a policy signal, but independent testing will be needed to understand both security capability and false refusals on legitimate defensive work.

Pricing also illustrates why model comparisons need live operational data. Tencent’s stated list price is roughly $0.90 per million input tokens, $2.70 per million output tokens, and $0.045 per million cached input tokens. OpenRouter currently lists Tencent Cloud routing at about $0.417 per million input and $1.251 per million output. Teams copying launch-post prices into a spreadsheet may be wrong before the review meeting begins. Pin the provider during evaluation, record the actual route and cache behavior, and compare cost per completed task rather than cost per raw token.

What to test before changing the default route

Teams already using Hy4 Preview should run the production model beside it on 50 to 100 production-shaped tasks. Keep prompts, provider, tool definitions and retry policy fixed. Log tool-call arguments, retries, latency, token consumption and human corrections. Separate text generation from visual and computer-use tests; a blended average can hide a regression in the exact capability your application depends on.

Long context deserves particular suspicion. A one-million-token window is a capacity claim, not a guarantee that the model retrieves the right fact from the middle, follows an instruction from the beginning, or maintains stable tool state across a long session. Test increasing context lengths with planted facts, conflicting updates and irrelevant material. Measure accuracy and latency together. The cheapest long context is the context you do not send, so retrieval and summarization architecture still matter even when the limit looks effectively infinite.

Keep the preview endpoint available as a rollback until Hy4 survives real traffic, and avoid changing model, provider and agent policy in the same deployment. If all three move at once, a production incident becomes an archaeology project. The fast migration visible in OpenRouter is an argument for modular routing and strong observability, not for treating every new production label as an automatic upgrade.

Community reaction reflects the same tension. The Hacker News launch discussion reached 291 points and 139 comments during the research window. Some practitioners welcomed Tencent’s explicit discussion of limitations and viewed open models as already useful production tools; others challenged the benchmark framing and the hardware reality of a 770-billion-parameter system. Both reactions are reasonable. Hy4 can be meaningfully open, commercially accessible and operationally impractical for a small team to host—all at once.

The chart’s real message is not that Tencent has won a benchmark race. It is that interchangeable APIs and model routers now let a release absorb trillions of tokens before the industry has finished reading the model card. OpenRouter usage is therefore a strong launch-day smoke test: developers found the endpoint credible enough to try at scale. It is not an acceptance test. The teams that benefit will be the ones that can migrate quickly while still demanding evidence from their own workloads.

Hy4’s No. 10 debut is worth attention because adoption is a fact and supremacy is still a claim. Benchmark it now, preserve the rollback, and let production traces—not a vendor table or a weekly rank—decide whether it earns the default route.

Sources: Tencent launch announcement, OpenRouter Hy4 model page, Tencent Hy4 Preview repository and model card, Hacker News discussion