Tencent Hy4 preview
Also known as: Tencent Hy4 / HY-4 / Tencent Hunyuan 4 / Hunyuan 4 / tencent-hy4
Tencent Hy's August 28, 2026 open-weight flagship reasoning model: a 770B-A49B MoE with a 1,048,576-token context window, high and no_think reasoning modes, and a focus on long-horizon coding, office analysis, game development, and scientific research.
Data sourced primarily from official releases (GitHub, Hugging Face, papers), then benchmark leaderboards, then third-party evaluators. Learn about our data methodology
| Type | Condition | Input | Output |
|---|---|---|---|
| Text | - | ¥6.00/ 1M | ¥18.00/ 1M |
| Type | Condition | Input | Output |
|---|---|---|---|
| Text | - | $0.834/ 1M | $2.50/ 1M |
| Type | TTL | Write | Read |
|---|---|---|---|
| Text | - | — | $0.042/ 1M |
“—” means the modality is not billed in that direction, or the vendor has not published a price for it.
Hy4 preview currently shows benchmark results led by HLE (15 / 185, score 55.40), SWE-Bench Pro - Public (6 / 59, score 65.70), GPQA Diamond (24 / 226, score 92.30). This page also consolidates core specs, context limits, and API pricing so you can evaluate the model from benchmark results and deployment constraints together.
Hy4 preview is Tencent Hy's new flagship reasoning model, released with open weights on August 28, 2026. It is the official release behind the earlier Tencent Hy4, HY-4, and Hunyuan 4 reports. Tencent describes it as an early Hy4 preview built for real productivity, with a focus on long-horizon software engineering, multi-file office analysis, game development, and scientific research.
The backbone is a Mixture-of-Experts model with 770B total parameters and 49B active parameters per token. Of its 78 backbone layers, the first uses a dense FFN and the other 77 each contain 256 routed experts plus one shared expert. Every token activates eight routed experts and the shared expert. A native MTP layer adds 10B total and 0.7B active parameters for speculative decoding.
Hy4 preview uses Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse-index reuse, together with four residual streams based on identity Hyper-Connections. The official configuration specifies a hidden size of 6,144, 64 attention heads, a 1,048,576-token context window, and a vocabulary of 120,832. Its primary input and output modality is text.
The model defaults to high reasoning and also supports no_think for direct responses. Tencent recommends temperature 0.9 and top_p 1.0. The open chat template includes structured tool-call and tool-response formats. Official vLLM and SGLang examples provide tool parsing, reasoning parsing, and MTP speculative decoding, using hy4-preview as the served model name.
Tencent released both the original BF16 weights and an FP8 quantized variant. The BF16 Hugging Face repository is about 1.42 TiB across 131 safetensors shards, so self-hosting generally requires a multi-GPU server. The code and model weights are licensed under Apache License 2.0.
Tencent's benchmark appendix covers agentic coding, search, workplace tools, STEM agents, and general reasoning. Representative Hy4 preview results include 82.9 on SWE-bench Multilingual, 65.7 on SWE-bench Pro, 64.3 on DeepSWE, 85.4 on Terminal-Bench 2.1, 78.4 on CyberGym, 83.7 on MCP-Atlas, 74.1 on Toolathlon-Verified, and 92.3 on GPQA Diamond. HLE scores are 43.4 without tools and 55.4 with tools. DataLearner adds structured rows only when a public benchmark and its tool setting can be matched precisely; Tencent's internal benchmarks are not mixed into public leaderboards.
In a separate blind comparison, 163 Tencent experts rated outputs on 203 engineering tasks. Hy4 preview averaged 2.99 out of 4, slightly above the same evaluation's GLM 5.3 score of 2.92 and Kimi K3 score of 2.94. These are vendor-run results, and real outcomes can vary with the agent scaffold, tools, timeout, sampling, and context setup.
The official API model ID is hy4-preview, available through Tencent Cloud TokenHub. Tencent's English release page lists prices per one million tokens at $0.042 for cached input, $0.834 for regular input, and $2.501 for output. The Chinese release page lists the corresponding domestic prices at CNY 0.3, CNY 6, and CNY 18. The model is also available in Tencent Yuanbao, ima, CodeBuddy, and WorkBuddy.
Tencent explicitly calls this an early Hy4 version with room for further pre-training and post-training improvements. Known issues include spending longer than necessary on complex reasoning and a tendency to repeatedly check or over-verify its own work. Preview status should therefore not be read as a claim of final-release stability.
Yes. Tencent released Hy4 preview on August 28, 2026 with both original BF16 and FP8 weights. Its code and model weights use Apache License 2.0, although Tencent still describes it as an early preview of Hy4.
The backbone has 770B total parameters and activates 49B per token. A native MTP layer adds 10B total and 0.7B active parameters. The official configuration supports 1,048,576 context tokens.
The official chat template supports high and no_think. High is the default for math, coding, and complex reasoning, while no_think produces more direct answers. Tencent recommends temperature 0.9 and top_p 1.0.
Tencent's international pricing per one million tokens is $0.042 for cached input, $0.834 for regular input, and $2.501 for output. The domestic Chinese prices are CNY 0.3, CNY 6, and CNY 18 respectively.
Tencent says the early model can spend longer than necessary on complex reasoning and may repeatedly check or over-verify its work. Its published benchmark figures are vendor-run and can vary with tools, agent scaffolds, timeouts, and sampling settings.
Follow DataLearner on WeChat for AI model updates and research notes.
