Tencent Hy4 Preview Reality Check: Agent Arena +6.92% at $0.26/Task, 1M Context and Benchmark Limits

Tencent's August 28 Hy4 preview release now has something the original launch article did not: several days of third-party Arena evidence. That makes a materially updated assessment possible without pretending the model's vendor benchmark chart and live user-agent measurements are the same thing.

What Hy4 preview actually is

Tencent's official repository describes Hy4 preview as a 770-billion-parameter Mixture-of-Experts model with 49 billion backbone parameters activated per token. It has 78 backbone layers, 256 routed experts plus one shared expert in each of the 77 MoE layers, top-8 routed-expert activation, a 1M-token context window and a native multi-token-prediction layer intended for speculative decoding.

The weights are genuinely available rather than merely promised. Tencent publishes Hy4 preview and an FP8 variant and licenses the project under Apache 2.0. The official deployment examples target vLLM and SGLang and expose reasoning and tool-call parsing through an OpenAI-compatible serving path. Tencent still labels this an early Hy4 version and explicitly lists over-reasoning and over-verification among known issues.

New independent signal: Agent Arena

Arena's September 5 Agent Arena gives a more useful outside signal than parameter count alone. Its overall board covers 2,285,256 agent sessions across 59 models. Hy4 preview sits on the cost/performance Pareto frontier with a reported net-improvement signal of +6.92% and a median cost of about $0.26 per task.

The code-specific view is also notable. On Arena's September 5 Code Agent view, which covers 686,333 sessions, Hy4 preview is on the Pareto frontier at +8.48% net improvement and about $0.47 per task. The same page reports a 15.91% confirmed-success signal for Hy4 preview and a 11.63% bash-recovery signal.

These are not pass rates from a static coding benchmark. Arena describes Agent Arena as a dynamic ranking built from real agent sessions and behavioral signals such as task confirmation, complaints/praise, steerability, bash recovery and tool hallucination. Cost per task is route- and workload-dependent. The numbers therefore tell us something about observed agent usefulness and economics in Arena's environment, not an intrinsic universal success rate for the weights.

Web-development evidence is competitive, not frontier-leading

Arena's Code Arena: WebDev page provides a separate pairwise preference signal. On the September 5 board, Hy4 preview is listed at 1621 with a displayed +16/-16 uncertainty and 1,661 model votes. That places it behind the newest proprietary frontier entries on that particular date, including GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.

This is a useful correction to broad claims that a strong open-weight coding benchmark automatically means "best coder." WebDev Arena measures interactive web-development preference, not repository issue resolution, terminal autonomy or general reasoning. Its result is also dynamic as votes and models change.

Tencent's internal productivity evaluation remains vendor evidence

Tencent's own blind engineering evaluation used 163 internal experts across 203 engineering tasks. The company reports Hy4 preview at 2.99/4.00, compared with 2.92 for GLM 5.3 and 2.94 for Kimi K3. Against GLM 5.3, Tencent reports 46.8% wins, 12.8% ties and 40.4% losses; against Kimi K3 it reports 51.2% wins, 7.9% ties and 40.9% losses.

The sample size and win/tie/loss breakdown are useful, but this remains a Tencent-designed task set evaluated by Tencent's own experts. It should not be merged with Arena's independent session signals into one synthetic leaderboard score.

Tencent also says Hy4 preview participated in optimizing its own inference system and that the resulting operator-fusion and communication changes increased end-to-end throughput by 31.8% over a baseline. That is a vendor engineering claim about an optimized inference stack, not evidence that the model generates 31.8% more output tokens per second on every provider or every hardware configuration.

SWE-bench Pro and SWE-bench Verified must stay separate

Tencent's official README places its detailed benchmark appendix in an image rather than machine-readable text. Public transcriptions of that vendor appendix consistently report a 65.7 result for Hy4 preview on SWE-bench Pro and 82.9 on SWE-bench Multilingual, alongside other agent/coding metrics. Those values should be treated as Tencent-reported launch-chart results unless an independent evaluator reproduces them with the same benchmark revision, harness, tools, reasoning setting and attempt budget.

SWE-bench Verified is a different benchmark family. In the first-party Tencent release text, repository material and current Arena pages checked for this update, no exact Hy4 preview SWE-bench Verified score with reproducible methodology was found. It therefore remains unknown here rather than being filled from SWE-bench Pro, SWE-bench Multilingual, DeepSWE or a generic "SWE-bench" label.

That distinction is especially important in 2026, when software-engineering scores can vary dramatically with scaffold, tool access, retry policy, reasoning budget and benchmark version.

Pricing and self-hosting economics

Tencent's corporate release lists hosted API pricing of $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cache-hit tokens. Arena's Agent pages display roughly $0.83/$2.50 and estimate about $0.26 per overall agent task or $0.47 per code-agent task in its own workload.

Those figures are consistent enough to make Hy4 preview economically interesting for hosted agent work, but "cheap per token" and "cheap per solved task" are not equivalent. Long reasoning, failed tool calls, retries and large contexts can dominate task cost.

Self-hosting removes per-token API billing but not infrastructure cost. A 770B-total MoE remains a large distributed-serving workload even with only 49B backbone parameters activated per token. Tencent's own example uses tensor parallelism across eight GPUs for the FP8 checkpoint, so open weights should not be read as laptop-scale deployment.

Early public feedback: interest, but not a consensus

A launch-week OpenCode discussion shows the predictable mix of excitement and uncertainty. Some commenters were enthusiastic about the model's pace and open availability, while others questioned comparator choices and asked whether Hy4 was actually better than DeepSeek, Muse Spark or other frontier options in long-context understanding. The thread itself concludes that it was still too early to know.

That is the appropriate evidentiary weight: user comments can identify questions worth testing, but they are self-selected anecdotes without controlled prompts, pinned model revisions, fixed routes or objective grading. No public consensus is inferred.

Practical takeaway

Hy4 preview now has a stronger evidence stack than it had at launch. The model is genuinely open-weight under Apache 2.0, has a 1M context window and production-oriented vLLM/SGLang recipes. Tencent's internal engineering evaluation and benchmark appendix suggest a large generational jump, while Arena now provides outside evidence that Hy4 can sit on the agent cost/performance Pareto frontier and remain competitive in interactive coding.

The limits are just as important. Arena's metrics are dynamic behavioral signals, not SWE-bench scores. Tencent's benchmark appendix is vendor evidence until matched independent reruns appear. SWE-bench Pro and SWE-bench Verified must not be conflated, and no reproducible Hy4 SWE-bench Verified result was found in the primary sources checked for this update.

For teams evaluating Hy4, the most defensible next step is a pinned workload test: exact checkpoint, exact serving stack, fixed reasoning settings, identical tools, measured TTFT and output speed, total tokens, task success, retries and cost per successful task. That will tell you more than any single launch chart or live leaderboard rank.