Hy4 preview (Deep Infra)

Tencent released and open-sourced Hy4 preview on August 28, 2026, announcing a next-generation large language model with 770B total parameters, 49B active parameters, and a context window exceeding 1M tokens. The model is available open-source and accessible through Tencent products (WorkBuddy, CodeBuddy, Yuanbao, ima) In an internal Tencent blind evaluation involving 163 experts and 203 engineering tasks, Hy4 preview scored an average of 2.99 out of 4.00, slightly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). The model was developed using high-quality training data co-created with Tencent experts across software engineering, gaming, f

Hy4 preview is Tencent's open-weight Mixture-of-Experts flagship language model released on August 28, 2026, with 770B total parameters, 49B activated per token, and a 1M-token context window. The 78-layer backbone uses 256 routed experts plus a shared expert in most layers, selects eight routed experts per token, and The 1M-token context window is positioned as enabling large repositories, document collections, extensive logs, and long-running agent sessions, though the article notes it does not guarantee perfect recall and increases input cost. Tencent positions software engineering as a central workload, with targeted capabilitie

Tencent released and open-sourced Hy4 preview on August 28, 2026 as the latest generation of its Hunyuan large language model family, with core specifications of 770B total parameters, 49B activated parameters, and a 1M-token context length — entering the ultra-large-scale tier of open-source models. The sparse activat Hy4 preview is positioned as "Built for productivity," connecting requirements comprehension, document ingestion, execution, validation, and delivery into longer end-to-end workflows rather than one-off tasks. The article argues the real test of the 1M-token context window is whether the model can locate relevant infor

Tencent open-sourced Hy4 preview on August 28, 2026 with Apache 2.0 weights (no field-of-use clause or geo carve-out), 770B total/49B active MoE parameters, a 1M-token context window (1,048,576 tokens) with 64K max completion, and an international price point of $0.834/M input, $2.501/M output, $0.042/M cached — cheape Architecture details reveal a 78-layer backbone (layer 1 dense FFN, layers 2-78 MoE), 256 routed experts plus 1 shared per MoE layer with top-8 routing, hidden size 6,144, 64 attention heads using Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse, Identity Hyper-Connections residual sch

Tencent Hunyuan released Hy4 preview as an open-source flagship model on August 28, 2026, featuring 770B total parameters, 49B activated parameters, and a 1M-token context length, simultaneously open-sourced on HuggingFace, GitHub, and ModelScope and available via Tencent Cloud TokenHub and OpenRouter. The model target Capability gains are reported across four directions: enhanced long-horizon development understanding/planning/debugging for software engineering; end-to-end delivery from data analysis to documents/tables/presentations for office scenarios; Unreal5 and Unity engine integration via MCP enabling playable demo generation

Tencent's Hy team officially announced Hy4 preview on August 28, 2026, releasing it as their most capable model to date. The model is a 770B-parameter Mixture-of-Experts system with 49B active parameters per token and a 1M-token context window. Tencent describes it as the largest generation-over-generation capability g The release is explicitly positioned for productivity workloads built with input from Tencent internal experts in software engineering, game development, finance, and security. Highlighted capability areas include long-horizon software engineering (understanding, planning, debugging, verification, plus front-end visual