Hy4 preview

IntuitionLabs' analytical write-up confirms Tencent released Hy4 preview on August 28, 2026 as an open-weight Mixture-of-Experts language model with 770B total parameters and 49B activated per token, citing Tencent's announcement, the Hugging Face model card, and Reuters. The model is released under Apache 2.0 with bot A key contribution of this candidate is its explicit caveat that Tencent's comparative benchmarks were conducted internally using a self-selected test suite, meaning the model's ranking should be read as a snapshot of that internal evaluation rather than a settled third-party-audited result. This methodology flag adds

TechNode Global reports that on August 28, 2026, Tencent released Hy4 preview as an open-source Mixture-of-Experts model with 770 billion total parameters, 49 billion activated per token, and a context window of more than 1 million tokens. The excerpt confirms the Hugging Face model card specifies a one-million-token c The excerpt details Hy4 preview's architecture: 78 backbone layers, most using a MoE structure with 256 routed experts and one shared expert, with eight routed experts plus the shared expert activated per token, plus a separate multi-token prediction layer for speculative decoding. Tencent also published instructions f

The miraflow.ai breakdown of August 29, 2026 reports that Tencent open-sourced Hy4 preview on August 28, 2026, and highlights a self-referential design detail: the model was used during its own development to help automate optimization of training methods, and it separately optimized its own inference infrastructure to The same article positions Hy4 preview within the Hunyuan lineage and walks through the model's MoE architecture, the Gated DeepSeek Sparse Attention mechanism, the four residual streams using Hyper-Connections, and the native multi-token prediction layer used for speculative decoding, while also comparing its benchmar

Tencent announced on August 28, 2026 the release and open-sourcing of Tencent Hy4 preview, a next-generation large language model with 770B total parameters and 49B active parameters, and a context window exceeding 1M tokens. According to the page, the model demonstrates strong capabilities on real-world productivity t The Tencent announcement also reports the methodology and results of an internal blind evaluation in which 163 experts rated Hy4 preview on 203 engineering tasks, producing an average score of 2.99 out of 4.00, slightly ahead of GLM-5.3 (2.92/4.00) and Kimi K3 (2.94/4.00). The page states that Hy4 preview was developed

MyClaw's release-day summary confirms Hy4 preview as Tencent's open-weight Mixture-of-Experts flagship with 770B total parameters, 49B activated per token, and a 1M-token context window, released on August 28, 2026. The write-up details the 78-layer backbone architecture using 256 routed experts plus a shared expert se The candidate notes the model is infrastructure-heavy and intended for multi-GPU inference, with access options including Tencent Cloud TokenHub, OpenRouter, or self-hosting of the released weights including the FP8 variant. A practical caveat from Tencent warns that the preview can spend too long reasoning and over-ve

The llm-stats.com directory page for Hy4 preview states the model has 770 billion parameters, uses a Mixture-of-Experts architecture in the frontier parameter tier (200B+), and ships under Apache License 2.0 allowing commercial use. The excerpt lists the model as released on August 28, 2026 by Tencent and identifies th The excerpt lists common comparison models including Gemini 3 Pro, Claude Mythos Preview, Qwen3.7 Max, Qwen3.8 Max, Claude Opus 4.8, and Muse Spark 1.1, with comparison scores marked as pending for each. It notes the model is open-weight and can be downloaded and self-hosted on the user's own hardware. The page serves

The Hugging Face model card for tencent/Hy4-preview provides the primary technical specification for the Hy4 preview model, confirming it as a Mixture-of-Experts design from the Tencent Hy Team with 770B total parameters and 49B activated per token. The 78-layer backbone uses one standard dense FFN layer followed by 77 The model card lists a 1M-token context length, vocabulary size of 120,832, and Apache 2.0 licensing, with full-precision and FP8 checkpoints distributed through Hugging Face alongside mirrors on ModelScope, CNB, and GitCode. Deployment quickstart sections cover vLLM, SGLang, finetuning, and quantization. The document