Hy4 preview

IntuitionLabs' September 5, 2026 analysis confirms Hy4 preview as a Mixture-of-Experts model released by the Tencent Hy Team on August 28, 2026, with 770B total parameters and 49B activated per token, released under Apache License 2.0 with weights shipped in both standard and FP8-quantized checkpoints on Hugging Face, The article notes that Tencent's published benchmark scores — including a 2.99/4.00 average over 203 engineering tasks across 163 internal experts that slightly edged GLM-5.3 (2.92) and Kimi K3 (2.94) — come from Tencent's own internal blind evaluation rather than a third-party-audited ranking, so the model's standing

TechNode's September 4, 2026 report confirms Hy4 preview as a Mixture-of-Experts model with 770 billion total parameters, 49 billion activated per token, and a context window of more than 1 million tokens, released under Apache License 2.0. The model uses 78 backbone layers—most with an MoE structure of 256 routed expe For developers, Tencent published both the full model and a lower-precision FP8 checkpoint, and released instructions for serving the model with vLLM and SGLang, along with fine-tuning guidance and an AngelSlim compression toolkit. These resources enable self-hosting beyond Tencent's hosted products, though operators m

FelloAI's September 1, 2026 coverage frames Hy4 preview — released and open-sourced on August 28, 2026 — as the largest permissively licensed model anyone has shipped, with 770B total / 49B active parameters, a 1,048,576-token context, 78 layers, 256 routed experts plus one shared, top-8 routing, and Apache 2.0 licensi The article reiterates Tencent's own benchmark numbers (Terminal-Bench 2.1 at 85.4, SWE-Bench Pro 51.2, DeepSWE 64.3, Toolathlon-Verified 74.1, APEX-Agents 37.1 pass@1) as Tencent-reported rather than third-party-verified, notes the absence of an Artificial Analysis rating, and states the single Arena placement is auto

ArWriterAI's August 31, 2026 explainer treats Hy4 preview as the fourth generation of Tencent's Hy (Hunyuan) model family, open-weighted on Hugging Face under Apache 2.0, with 770B total parameters / 49B activated per token, a 1M-token context, 78 transformer layers with 257 experts (256 routed + 1 shared, top-8 routin The article emphasizes the Apache 2.0 commercial-permissive license as the key differentiator versus Kimi K3's custom licence and GLM 5.3's bespoke Z.ai document, and highlights practical implications for content operations, including better handling of long brand-guideline stacks and long-form research corpora within

Developers Digest's August 31, 2026 piece corroborates that Tencent's Hy4 preview ships with 770B total parameters and 49B active per token under Apache 2.0, framed as a 1M-context text-only MoE with DeepSeek-style sparse attention. The article reports Terminal-Bench at 85.4 and DeepSWE at 64.3, identified as figures s The piece links to the official Hy4-preview model card on Hugging Face for architecture, benchmark appendix, quickstart, and deployment recipes, and references Simon Willison's independent reasoning-mode analysis alongside the Gated DeepSeek Sparse Attention and IndexCache papers cited by the model card. OpenRouter pri

Tencent's press release of August 28, 2026 confirms the open-source release of Hy4 preview as a next-generation large language model with 770B total parameters, 49B active parameters, and a context window exceeding 1M tokens, positioning it among the top tier of open-source models for real-world productivity across cod The release frames Hy4 preview as a major leap from gains in both pre-training and post-training and notes it was developed in deep co-design with products like CodeBuddy and WorkBuddy using training data co-created with Tencent experts in software engineering, gaming, finance, and security. It reports results from Ten

Aimodelsnavi's profile of Tencent Hy4 preview lists the model as Tencent's next-generation open-source flagship MoE, released August 28, 2026, with 770B total / 49B active parameters, a 1M-token (1,048,576) context window, and Apache 2.0 licensing. It details a 78-layer MoE design with 256 routed experts plus one share The same profile lists API pricing at $0.834 input / $2.501 output per 1M tokens with cache-hit input at $0.042/1M, weights shipped on Hugging Face and ModelScope, and API access via Tencent Cloud TokenHub and OpenRouter. It frames Hy4 preview as a productivity-focused model for long-horizon software engineering, compl

Tencent's official research page announces the release of Hy4 preview on August 28, 2026, describing it as the company's most capable model to date: a 770B-parameter Mixture-of-Experts design with 49B active parameters per token and a 1M-token context window. The page frames the release as a flagship generation built a The page highlights a productivity-focused co-design with Tencent internal experts in software engineering, game development, finance analysis, and security, with worked examples that span a Three.js museum-style 3D viewer of seven Chinese historic buildings, a reimbursement-claim review that reasons across policies, r