AI/TLDR

Hunyuan Hy4 preview

Tencent's 770B-parameter open-weight Mixture-of-Experts preview with a 1M-token context, released August 28, 2026 under Apache 2.0.

Overview

Hunyuan Hy4 preview is a 770B-parameter Mixture-of-Experts language model from the Tencent Hy Team, released and open-sourced on 28 August 2026. It activates 49 billion parameters per token: the backbone runs 78 layers, of which the first uses a dense FFN and the remaining 77 are MoE layers holding 256 routed experts plus 1 shared expert, with the top-8 routed experts and the shared expert firing on every token. A native Multi-Token Prediction layer (10B total, 0.7B activated) is built in for speculative decoding. The context length is 1M tokens and the vocabulary is 120,832 entries.

On the architecture side, Tencent credits DeepSeek and GLM as inspirations: attention uses Gated DeepSeek Sparse Attention (Gated DSA) with IndexCache for cross-layer sparse-index reuse, and the residual pathway uses identity Hyper-Connections (iHC) with four residual streams to widen inter-layer information flow. Weights ship in BF16 as tencent/Hy4-preview and in FP8 as tencent/Hy4-preview-FP8, both under the Apache 2.0 license, and Tencent publishes vLLM and SGLang serving recipes with prebuilt container images.

Tencent positions Hy4 preview as a productivity model, built with in-house software engineers, game developers, finance analysts and security experts on the work those teams ship. In a blind side-by-side evaluation, 163 internal experts rated model outputs on 203 engineering tasks: Hy4 preview averaged 2.99 against GLM 5.3 at 2.92 (46.8% wins / 12.8% ties / 40.4% losses) and against Kimi K3 at 2.94 (51.2% wins / 7.9% ties / 40.9% losses). Tencent also publishes the model's known limitations up front — it can spend longer than necessary reasoning through complex tasks and tends to over-verify its own work — and frames the release as a deliberate early ship, the same approach it took with Hy3 preview.

Released2026-08-28
LicenseApache-2.0
WeightsOpen weights
Parameters770B total · 49B active
Context1M
ArchitectureMixture-of-Experts (78 layers, 256 routed experts + 1 shared, top-8 activated) with Gated DeepSeek Sparse Attention
ModalitiesText
StatusPreview — Tencent describes Hy4 preview as an early version of the Hy4 generation. Weights are published on Hugging Face, ModelScope, GitCode and CNB; hosted access runs through Tencent Cloud TokenHub and OpenRouter.

Benchmarks

Hy4 preview vs Hy3 and named peers, transcribed from the benchmark appendix in Tencent's Hy4 preview model card. Tencent evaluates every model at its highest available reasoning setting and measured the competitor figures itself; where the appendix prints two figures for a cell, both are shown as a/b. A dash means the appendix left the cell blank.

BenchmarkHy3Hy4 previewDeepSeek V4 Pro 0813Qwen 3.8 MaxGLM 5.3Kimi K3GPT 5.6 SolClaude Opus 5
SWE-bench Multilingual75.8%82.9%77.3%82.6%81.3%80.8%74.1%89.5/85.8%
SWE-bench Pro57.9%65.7%60.3%67.7/61.6%64.6%63.3%64.6/60.5%79.2/79.9%
DeepSWE28%64.3%62.7/58.8%56.6/55.6%66.9/68.1%67.5/74.0%72.7/68.9%68.8/74.7%
SWE Atlas — Codebase Q&A30.8%64%53.4%55.4%55.8%35.2%58.1%57.5%
SWE Atlas — Test Writing35.9%57.8%45.6%52.8%49.6%35.6%49.6%68.7%
SWE Atlas — Refactoring32.9%53.3%48.6%51%51.9%37.4%52.4%60%
Terminal-Bench 2.170.8%85.4%87.9/80.3%86.6/85.8%88.2/88.3%88.3/85.7%88.8/88.3%86.7/85.4%
ProgramBench3%17.5%15.5%17.5%18%24.5%25%39.5%
Harbor-Index15.6%39.6%36.9%38.8%42.5%46.3%56.9%
WideSearch81.9%83.9%81.8%81.9/81.1%83.2%81%86.3%84%
OneMillionBench (with tools)51.5%65.4%62%63.1%64.5%63.5%67.1%68.1%
MCP-Atlas (public)75%83.7%82.5%81.9%81.9%84.2/82.8%82.5%85.7%
BankerToolBench68.8%78.6%73.1%74.7%77.8%73.5%79%81.9%
Agents' Last Exam (ALE-CLI)17.1%22.8%21.9%25.4%23.8%23.2%27.6%25.1%
GDPval-AA V2 (official)1213 Elo1678 Elo1580 Elo1717 Elo1763 Elo1675 Elo1711 Elo1831 Elo
BioMysteryBench54.9%71.3%61.6%58.9%69%61.3%73.1%72.1%
GPQA Diamond90.9%92.3%92.8/91.7%92.6/92.2%91.7/91.4%93.5/92.8%94.1/94.7%93.7/93.3%
SUPERChem52.6%66.4%62%61.9%58.5%66.9%73.6%76.7%
MathArena Apex 202538.7%74.2%66.3%72.8%68.4%90%91.4%

Comparison source ↗

This model's scores

  1. GPQA Diamond92.3%
  2. Terminal-Bench 2.185.4%
  3. MCP-Atlas (public)83.7%
  4. SWE-bench Multilingual82.9%
  5. MathArena Apex 202574.2%
  6. SWE-bench Pro65.7%
  7. DeepSWE64.3%
  8. HLE (no tools, text-only)43.4%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

Pricing

Input$0.834 / 1M tokens
Cached input$0.042 / 1M tokens
Output$2.501 / 1M tokens

API pricing quoted in Tencent's launch announcement; served through Tencent Cloud TokenHub and OpenRouter. Weights are Apache 2.0, so self-hosting is free.

Pricing source ↗

Strengths

  • Apache 2.0 open weights at 770B total parameters (49B active), free for commercial, self-hosted and on-premise use
  • 1M-token context window with Gated DeepSeek Sparse Attention and IndexCache sparse-index reuse
  • Native Multi-Token Prediction layer for speculative decoding, wired into both the vLLM and SGLang recipes
  • Published head-to-head appendix against Hy3, DeepSeek V4 Pro, Qwen 3.8 Max, GLM 5.3, Kimi K3, GPT 5.6 Sol and Claude Opus 5
  • Large generation-over-generation jumps on agentic coding — DeepSWE 28.0 → 64.3 and Terminal-Bench 2.1 70.8 → 85.4 over Hy3
  • BF16 and FP8 checkpoints distributed across Hugging Face, ModelScope, GitCode and CNB

Best for

  • Self-hosted agentic coding and long-horizon software engineering via vLLM or SGLang
  • Office and analysis work that turns context spread across many files into documents, spreadsheets and presentations
  • Long-context codebase and document analysis that needs the full 1M-token window
  • On-prem enterprise deployments where an Apache-2.0 licence is a requirement
  • Game-development prototyping and scientific-research reasoning workloads

How to access

ProviderModel ID
Tencent Cloud TokenHub ↗hy4-preview
OpenRouter ↗tencent/hy4-preview
Hugging Face (self-hosted via vLLM / SGLang) ↗tencent/Hy4-preview

Hunyuan — every version

The full lineage of the Hunyuan line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.

VersionReleasedContextLicense
Hunyuan Hy4 preview2026-08-281MApache-2.0
Hunyuan Hy3current2026-07-06256KApache-2.0

FAQ

Is Hunyuan Hy4 preview open source?

Yes. Tencent published Hy4 preview (BF16) and Hy4 preview-FP8 under the Apache 2.0 license on Hugging Face, ModelScope, GitCode and CNB, so the weights can be downloaded, fine-tuned and used commercially, including on-premise. Tencent also ships a finetuning pipeline with the model card and points at AngelSlim for compression and quantization.

How large is Hy4 preview and how much context does it handle?

It is a Mixture-of-Experts model with 770 billion total parameters, 49 billion of which activate per token. The backbone is 78 layers deep — the first uses a dense FFN and the other 77 are MoE layers with 256 routed experts and 1 shared expert, top-8 routed experts firing per token — plus a native Multi-Token Prediction layer of 10B parameters (0.7B activated) for speculative decoding. Context length is 1M tokens.

How does Hy4 preview compare with Hy3?

Tencent's published appendix shows large generation-over-generation gains on agentic work: DeepSWE goes from 28.0 to 64.3, Terminal-Bench 2.1 from 70.8 to 85.4, SWE-bench Pro from 57.9 to 65.7, MathArena Apex 2025 from 38.7 to 74.2, and the official GDPval-AA V2 Elo from 1213 to 1678. Reasoning moves less: GPQA Diamond goes from 90.9 to 92.3.

Why is it called a preview?

Tencent ships it as an early version of the Hy4 generation and says there is real headroom left in both pre-training and post-training. It documents two known issues at launch — the model can spend longer than necessary reasoning through complex tasks, and it tends to over-verify its own work. Tencent frames this as the same ship-early approach it used with Hy3 preview, which it credits for making the final Hy3 substantially better.

How do I run Hy4 preview?

Tencent recommends vLLM or SGLang and publishes prebuilt images for both. For vLLM: `docker run --gpus all -p 8000:8000 --ipc=host vllm/vllm-openai:hy4-preview tencent/Hy4-preview-FP8 --tensor-parallel-size 8 --attention-backend FLASHMLA_SPARSE --tool-call-parser hy_v4 --reasoning-parser hy_v4 --enable-auto-tool-choice`. Reasoning defaults to `high`; pass `chat_template_kwargs.reasoning_effort = "no_think"` for direct answers. Recommended sampling is temperature 0.9, top_p 1.0.