Benchmark Hy4 preview Hy3 Qwen 3.8 Max DeepSeek V4 Pro 0813 GPT 5.6 Sol Z GLM 5.3 Kimi K3 Claude Opus 5
Terminal Bench 2.1 85.4 70.8 80.3 85.4 85.7 85.8 88.3 88.3
DeepSWE 64.3 28.0 55.6 58.8 68.1 68.9 74.0 74.7
ProgramBench 17.5 3.0 15.5 17.5 18.0 24.5 25.0 39.5
SWE Atlas Refactoring 53.3 32.9 37.4 48.6 51.0 51.9 52.4 60.0
Agents' Last Exam (ALE-CLI) 65.4 51.5 62.0 63.1 63.5 64.5 67.1 68.1
Toolathon-Verified 71.3 54.9 58.9 61.3 61.6 69.0 72.1 73.1
APEX-Agents (pass@1) 43.4 34.4 40.5 41.5 42.3 46.6 49.6 53.2
PostTrainBench 8.8 3.5 4.4 5.3 5.3 7.1 10.6 —

On August 28, 2026, Tencent's Hunyuan team released and open-sourced Hy4 preview, a new-generation Mixture-of-Experts flagship model. The model has 770 billion total parameters and 49 billion activated parameters per token. It supports a context window exceeding 1 million tokens.

Hy4 preview is available through Hugging Face, ModelScope, GitCode, CNB, Tencent Cloud TokenHub, OpenRouter, and Tencent products including Yuanbao, ima, WorkBuddy, and CodeBuddy.

What's new

  • Architecture: MoE, 770B total / 49B activated parameters, 78 layers, 6144 hidden size, 64 attention heads
  • Attention: Gated DeepSeek Sparse Attention with IndexCache (32 indexer heads, 128 head dimension, top-k 2048)
  • Residual: 4 identity Hyper-Connection streams
  • Experts: 256 routed + 1 shared per MoE layer, top-8 activated per token
  • Context: 1M tokens, vocabulary 120,832
  • License: Apache 2.0
  • Deployment: vLLM and SGLang recipes with FP8 quantization and speculative decoding via MTP
  • API pricing: $0.834 per million input tokens, $2.501 per million output tokens, $0.042 per million cache-hit tokens
  • Free access: WorkBuddy and CodeBuddy free for two weeks; Hy3 free extended to September 30

Productivity benchmarks and internal evaluation

Tencent ran a blind side-by-side evaluation with 163 internal experts across 203 engineering tasks. Hy4 preview scored an average of 2.99 out of 4.00, edging out GLM 5.3 (2.92) and Kimi K3 (2.94). Win rates were 46.8% versus GLM 5.3 with 12.8% ties, and 51.2% versus Kimi K3 with 7.9% ties.

The company says gains come from scaling model size, context length, and training data, plus deep co-design with domain experts in software engineering, game development, finance, and security.

In software engineering, the model improves long-horizon understanding, planning, debugging, and front-end visual quality. In office analytics, it handles messy multi-file context and produces documents, spreadsheets, and presentations with stronger financial modeling. In game development, it generates playable prototypes from a single prompt and iterates with game engines. In scientific research, it shows progress in AI R&D, molecular dynamics, condensed matter physics, and pure mathematics.

Self-improvement loop and inference optimization

Hy4 preview participated in its own development for the first time. It proposed training-method optimizations, ran experiments, and fed results back into subsequent rounds — an early recursive self-improvement loop. The model also autonomously analyzed bottlenecks in its inference system, optimizing operator fusion and communication. These changes increased end-to-end throughput by 31.8% versus the baseline across context lengths and concurrency levels.

Scientific milestones

The model contributed to several research advances. It collaborated with Hyra on machine-learning force-field molecular dynamics for material screening and drug development. It independently constructed a quantum scattering solver and robustly optimized a five-barrier structure in low-temperature quantum transport, significantly reducing average leakage in high-energy bandgaps. It also made progress on the three-dimensional Blaschke–Lebesgue problem, pushing the volume lower bound to 0.41104 — within 2% of the conjectured optimum.

Known limitations

Tencent acknowledges this is an early preview with headroom in both pre-training and post-training. Known issues include spending longer than necessary reasoning through complex tasks and a tendency to over-verify its own work. The team plans rapid iteration based on real-world feedback, following the same preview-first approach that substantially improved Hy3.

Why it matters

Hy4 preview pushes the open-source frontier toward 1M-token context at 770B scale while keeping activated parameters at 49B — a ratio that makes serving more tractable than dense equivalents. The Apache 2.0 license and immediate availability on multiple registries and inference stacks lower the barrier for organizations that want to self-host a frontier-class MoE. The self-reported 31.8% inference throughput gain from model-driven optimization also signals a shift: the model is beginning to improve the infrastructure it runs on.

Our take

Tencent's preview-first cadence — three major releases in six months — compresses the feedback loop between research and production. The 49B activated count keeps hardware requirements within reach of 8×H100-class nodes for FP8 serving, but the 770B total still demands serious VRAM and tensor-parallel orchestration. The real test will be whether reasoning-depth and over-verification issues resolve without sacrificing the coding and science gains that put Hy4 preview slightly ahead of GLM 5.3 and Kimi K3 in Tencent's blind eval.

Sources