Tencent releases 770B-parameter Hy4 Preview open weights

• Hy4 Preview ships as Apache 2.0 weights with a 770B-parameter MoE backbone, 49B activated parameters per token, and a 1M-token context claim, according to TencentHunyuan's launch announcement and the official model card.
• The 78-layer backbone uses 256 routed experts plus one shared expert, while the vLLM integration post describes cross-layer reuse of the model's sparse-attention index.
• Day-one access spans the AI Gateway announcement, the Cline announcement, and the OpenCode Go announcement, alongside Tencent's own WorkBuddy and CodeBuddy products.
• Tencent's reported SWE-bench Pro chart puts Hy4 at 65.7%, ahead of GPT 5.6 Sol and GLM 5.3 at 64.6%, as shown in a Cline post sharing Tencent's chart.

The Hugging Face model card includes a ready-made vLLM container, not just a checkpoint: it enables MTP speculative decoding, sparse attention, tool calling, and reasoning parsing with eight-way tensor parallelism. Tencent's known-limitations section says the preview can reason for too long and over-verify its work.

Tencent published both Hy4 Preview and an FP8 variant under Apache 2.0, with the weights and deployment material on Hugging Face and code in its GitHub repository. The model card says the 770B and 49B figures cover the backbone and exclude its native multi-token-prediction layer.

The backbone specification is:

• 78 layers and a 6,144-wide hidden state.
• One dense FFN first layer, followed by 77 MoE layers.
• 256 routed experts and one shared expert in each MoE layer.
• Top-8 routed experts activated for each token, alongside the shared expert.

The model card lists a 120,832-token vocabulary, 64 attention heads, and a 1M-token context length.

Hy4 uses Gated DeepSeek Sparse Attention with IndexCache, which the model card describes as cross-layer sparse-index reuse. The indexer selects a top-k of 2,048 tokens rather than applying dense attention across the full stated context.

vLLM's post gives the layer-level breakdown: 21 of 78 layers compute their own sparse index, while the other 57 reuse one. It also says the checkpoint includes a 10B-parameter native MTP layer, 0.7B active, with draft depth three for speculative decoding.

vLLM reported day-one NVIDIA validation for Hy4 and said its HPC-Ops attention and MoE kernels have been in vLLM mainline since Hy3 in its deployment post. Tencent's vLLM recipe uses the prebuilt vllm/vllm-openai:hy4-preview image and the FP8 checkpoint.

• --tensor-parallel-size 8 sets eight-way tensor parallelism.
• --speculative-config enables MTP with three speculative tokens.
• --attention-backend FLASHMLA_SPARSE selects the sparse-attention backend.
• The server enables Hy4 tool-call and reasoning parsers, plus automatic tool choice.

The same deployment page ships an lmsysorg/sglang:hy4-preview image for x86 and Arm, with its own three-step speculative configuration and tensor-parallel size of eight.

On SWE-bench Pro, a Cline post sharing Tencent's chart shows Tencent reporting Hy4 Preview at 65.7%, versus 64.6% for GPT 5.6 Sol and 64.6% for GLM 5.3.

Tencent also ran a blind side-by-side evaluation with 163 internal experts over 203 engineering tasks. Its model card reports a 2.99 average for Hy4 against 2.92 for GLM 5.3, with 46.8% wins, 12.8% ties, and 40.4% losses. Against Kimi K3, it reports 2.99 versus 2.94, with 51.2% wins, 7.9% ties, and 40.9% losses.

The benchmark chart and the expert evaluation are Tencent-supplied results.

Tencent says Hy4 is available globally through WorkBuddy, CodeBuddy, Yuanbao, and ima in its release announcement. The AI Gateway announcement names the hosted model ID as tencent/hy4-preview.

For coding-agent surfaces, the Cline announcement gives an npm i -g cline installation path followed by /model selection, while the OpenCode Go announcement says the preview is selectable there. A WorkBuddy preview post advertised a two-week free window and showed Hy4 in the model picker.

Tencent calls Hy4 an early version with remaining pre-training and post-training headroom. Its model card names two current behaviors: longer-than-needed reasoning on complex work and a tendency to over-verify results.

The default API mode is high reasoning, which the quickstart frames for math, coding, and reasoning tasks. Direct-response mode is exposed through chat_template_kwargs with reasoning_effort: "no_think".