Hy4 Preview

Hy4 Preview

Quick answer: Hy4 preview is Tencent Hunyuan's August 2026 flagship — a 770B-parameter Mixture-of-Experts model with 49B active parameters per token, a 1M-token context window, and Apache 2.0 open weights. It scores 92.3% on GPQA Diamond and 82.9% on SWE-bench Multilingual, a major generation-over-generation jump from Hy3.

At a Glance

Where Hy4 preview leads

• 92.3% GPQA Diamond — near-parity with the strongest closed models
• 82.9% SWE-bench Multilingual and 64.3% DeepSWE — large agentic-coding gains over Hy3 (which scored 75.8% and 28.0% respectively)
• 1M-token context window with Apache 2.0 open weights — fully self-hostable at frontier scale
• Native MTP (multi-token prediction) layer built in for speculative decoding

Where it lags

• Still trails the strongest proprietary models (GPT 5.6 Sol, Claude Opus 5) on several agentic and reasoning benchmarks
• Team-acknowledged issues: over-long reasoning traces and a tendency to over-verify its own work
• 770B total parameters (49B active) requires substantial multi-GPU infrastructure to self-host at full precision

Best for: Teams that need a frontier-class, fully open-weight (Apache 2.0) model for agentic coding, tool use, and long-context workloads without API lock-in.

What Hy4 Preview Is

Hy4 preview is the newest flagship in Tencent Hunyuan's "Hy" model line, succeeding Hy3 preview. It's a Mixture-of-Experts model with 78 backbone layers (the first dense, the remaining 77 MoE with 256 routed experts + 1 shared expert, top-8 routed experts active per token), plus a native multi-token-prediction (MTP) layer for speculative decoding. The architecture borrows ideas from DeepSeek and GLM: attention uses Gated DeepSeek Sparse Attention (Gated DSA) with IndexCache for cross-layer sparse index reuse, and the residual stream uses identity Hyper-Connections (iHC) to widen inter-layer information flow.

Tencent frames this release around agentic productivity rather than raw knowledge benchmarks alone — the team built training data directly from internal software engineers, game developers, finance analysts, and security experts, and reports the largest generation-over-generation gain they've measured, particularly in agentic coding (SWE-bench Multilingual jumped from Hy3's 75.8% to 82.9%; DeepSWE from 28.0% to 64.3%). In a blind internal side-by-side (163 experts rating 203 engineering tasks), Hy4 preview edged out both GLM 5.3 and Kimi K3.

This is an early "preview" release — Tencent explicitly flags known issues (over-long reasoning chains, excessive self-verification) and says they expect to iterate quickly, the same pattern that took Hy3 preview from launch to a substantially stronger model.

Specifications

Pricing

Open weights under Apache 2.0 — self-hosted via vLLM or SGLang (official prebuilt images vllm/vllm-openai:hy4-preview and lmsysorg/sglang:hy4-preview). No API list price; cost is your own inference infrastructure. An FP8-quantized variant (Hy4-preview-FP8) is also published for lower-memory deployment.

Public Benchmark Scores

Scores as self-reported by Tencent Hunyuan for Hy4 preview's own column in the model card's benchmark appendix table (some third-party comparison scores in that same table are Tencent's own re-tests). Several additional benchmarks in the model card (AutomationBench v1.0.6, PostTrainBench V1.1, SkillsBench 79-task text-only subset, Agents' Last Exam ALE-CLI, HLE with-tools variant, and several Tencent-internal-only evals) are not listed here — see tasks/hy4-preview-progress.md for why each was excluded (version/subset mismatch with the existing tracked benchmark, or no independent public benchmark page).

Hy4 Preview vs Alternatives

Comparison scores marked * are Tencent's own re-tests of the other vendors' models (from the same model-card appendix table), not those vendors' own self-reported numbers — see tasks/hy4-preview-progress.md for the full sourcing note.

Hy4 preview leads its own generation (Hy3) by a wide margin on both agentic coding and knowledge benchmarks, and is competitive with — though not consistently ahead of — the strongest proprietary models (GPT 5.6 Sol, Claude Opus 5) while being the only model in this set with fully open Apache 2.0 weights at this parameter scale.

Related Models & Benchmarks

• Compare: Hy3 · Kimi K3 · DeepSeek V4 Pro · GLM 5.3
• Benchmarks: GPQA Diamond · SWE-bench Multilingual · DeepSWE · Terminal-Bench 2.1

Frequently Asked Questions

Specs from Tencent Hunyuan's official Hy4-preview Hugging Face model card and GitHub repository (August 2026). Last updated 2026-08-31.