Hy4
Hy4 preview (tencent/hy4-preview) is Tencent Hy Team's open-weight, 770B-total / 49B-active Mixture-of-Experts flagship (Apache 2.0, released Aug 28, 2026) built for long-horizon coding agents, complex tool-use workflows, and productivity tasks, backed by a 1,048,576-token context window and DeepSeek-style sparse attention.
Empirical Evaluation Results
Architectural Profile & Capabilities
Hy4 preview (OpenRouter slug 'tencent/hy4-preview', Benchmark Heaven ID 'hy4-preview::default') is a new-generation Mixture-of-Experts (MoE) flagship model developed by the Tencent Hy Team and released on August 28, 2026 as open weights under Apache 2.0. The model comprises 770B total parameters of which ~49B are activated per token. Per the published variant-specific technical account, the backbone spans 78 layers (one dense feed-forward layer plus 77 MoE layers); each MoE layer contains 256 routed experts plus one shared expert, and each token activates top-8 routed experts plus the shared expert. Hidden size is reported at 6,144 with a 120,832-token vocabulary; attention is described as Gated DSA with IndexCache (a DeepSeek-style sparse-attention scheme that reuses sparse attention indices across layers), and the residual design uses four residual streams (iHC). An integrated speculative-decoding / MTP component is reported to push the included parameter count toward ~780B, while Tencent's headline 770B figure refers to the backbone. The model is text-only and exposes a 1,048,576-token (1M) context window. Tencent positions Hy4 explicitly for coding agents, multi-step tool-use pipelines, long-horizon agentic execution, and productivity workloads; availability spans Tencent Cloud TokenHub, OpenRouter, and Tencent first-party products (WorkBuddy, CodeBuddy, Yuanbao, ima). Pricing on OpenRouter is $0.7506/1M input and $2.2509/1M output. Tencent reports a 31.8% end-to-end throughput improvement over an unspecified baseline, and an internal blind evaluation (163 experts, 203 engineering tasks) in which Hy4 preview averaged 2.99/4.00 versus GLM-5.3 at 2.92 and Kimi K3 at 2.94. Verified variant-specific benchmarks beyond the retained Benchmark Heaven ground truth include SWE-bench Pro (65.7, public split), SWE-bench Multilingual (82.9%), DeepSWE (64.3), Toolathlon-Verified (74.1), APEX-Agents pass@1 (37.1), CyberBench v1.1 (65.4%), IOI (59.3%), LegalBench (83.8%), MedCode (43.2%), MedScribe (83.6%), ProofBench v1.1 (75.0%), Tax Agent Bench (32.6%), and SRE Bench (2.3%). Standard suites such as SWE-bench Verified, GPQA Diamond, AIME 2026, HLE, BrowseComp, OSWorld, MRCR v2, or an AA Intelligence Index score for this exact variant were not verified in the retrieved sources and are therefore omitted.
770B total parameters with ~49B active per token. 78 layers total: 1 dense FFN layer + 77 MoE layers. Each MoE layer holds 256 routed experts plus 1 shared expert; each token routes to top-8 routed experts plus the shared expert. Hidden size 6,144; vocabulary 120,832. Attention is Gated DSA with IndexCache (DeepSeek-style sparse attention reusing sparse indices across layers). Four residual streams (iHC). An integrated speculative-decoding/MTP head brings the included parameter count to ~780B while the 770B headline refers to the backbone. Text-only, 1,048,576-token context. Open weights under Apache 2.0.
Recommended Workloads & Primary Use Cases
- •Context economics: the full 1,048,576-token window is exposed, but effective recall at extreme lengths is unverified (no published MRCR v2 / long-context needle score for this variant) and very long prompts materially inflate latency and input cost at $0.7506/1M.
- •Provider variance: OpenRouter lists up to 4 serving providers; quantization, TTFT, and throughput (including speculative decoding / MTP availability) will differ by provider, so benchmark and latency figures are not uniform across deployments.
- •Weakness in operational/ops domains: SRE Bench scored only 2.3% and Tax Agent Bench 32.6%, indicating narrow generalization in some specialized agentic and finance tasks — validate before production reliance.
- •No verified scores on common academic/standard suites (SWE-bench Verified, GPQA Diamond, AIME 2026, HLE, BrowseComp, OSWorld, AA Intelligence Index) for this exact variant — treat these as unbenchmarked rather than assume capability.
- •Preview designation: this is a 'preview' release; behavior, routing, post-training, and pricing may shift, and some published headline scores (e.g. Terminal-Bench 2.1) diverge from independently calibrated figures — pin versions and re-validate.
- •Text-only modality: no vision/audio input, so multimodal agent pipelines require a separate model.
Calls to Hy4 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Tencent.
More Models from Tencent
Compare other engines available in this laboratory.