Hy4 preview

IntuitionLabs published a detailed analysis on September 5, 2026 of Tencent's Hy4 preview release, confirming the Mixture-of-Experts design with 770 billion total parameters and 49 billion activated per token, released under Apache License 2.0 on August 28, 2026. The article notes that weights are published in two form The IntuitionLabs piece explicitly cautions that Tencent's blind comparative evaluation was conducted internally, with detailed results and methodological limitations appearing in the analysis's "Evaluation Methodology and Limitations" section. According to the report, these caveats do not invalidate Hy4 preview's publ

Wavect's review, checked on September 2, 2026, restates that Tencent released Hy4 preview on August 28, 2026 with 770B backbone parameters, 49B active parameters per token, and a context window above one million tokens, with access routes via WorkBuddy, CodeBuddy, Tencent Cloud TokenHub, and OpenRouter, and two weeks o The review recommends a controlled coding-agent pilot for teams limited by long context, multi-step tool use, or repeated repository retrieval, while cautioning that self-hosting the 770B model is a datacenter-scale project and that Tencent labels this an early version that can reason and verify longer than necessary.

AI Profit Boardroom confirms Tencent open-sourced Hy4 preview on August 28, 2026 as a Hunyuan-family model with 770B total parameters, 49B active per token, and a context window beyond one million tokens, with two weeks of free access inside WorkBuddy and CodeBuddy at launch. The piece describes the MoE design as delivering frontier-scale capability at compute cost closer to a mid-sized model, enabling entire codebases or multi-day agent sessions within a single context, and positions productivity targets as coding, office tasks, data analysis, game development, and scientific research.

Developers Digest, verified against the model card on August 31, 2026, characterizes Hy4 preview as a 770B-total/49B-active text-only MoE with a 1M-token context window, Apache 2.0 license, and weights distributed on Hugging Face, ModelScope, GitCode, and CNB. It notes a 2.6x increase in total parameters and context le The review reports benchmark scores of Terminal-Bench 85.4 and DeepSWE 64.3, and lists DeepSeek-style sparse attention with IndexCache and the iHC residual pathway among the architectural details drawn from the model card.

Emergent.sh reports that Tencent launched Hy4 preview under the Apache 2.0 open-source license on August 28, 2026, positioning the model as an early-access preview that lets developers and researchers test capabilities and provide feedback before a full production release. The Apache 2.0 license is described as permitt The article frames Hy4 preview within Tencent's broader AI model portfolio alongside the Hunyuan series and specialized computer vision, NLP, and multimodal models, noting Tencent's accelerated 2026 AI development competing with domestic rivals Alibaba, Baidu, and ByteDance as well as international foundation-model pla

The Hugging Face model card for tencent/Hy4-preview is the first-party technical reference for the exact subject. It documents Hy4 preview as a Mixture-of-Experts flagship from the Tencent Hy Team with 770B total parameters and 49B activated per token, a 78-layer backbone where the first layer is a standard dense FFN a Because the page is the official model card, headline framing should be restricted to facts the excerpt itself states: Hy4 preview is a 770B/49B MoE with 1M context under Apache 2.0, using Gated DSA + IndexCache and iHC residuals with native MTP, and deployment recipes exist for vLLM and SGLang. Any specific benchmark

Tencent's official release confirms that Tencent Hy4 preview was released and open-sourced on August 28, 2026, as a next-generation large language model with 770B total parameters, 49B active parameters, and a context window exceeding 1M tokens, positioned for productivity tasks across coding, office work, and scientif The announcement reports access via WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub, and OpenRouter, with two weeks of free access on WorkBuddy and CodeBuddy at launch. In Tencent's internal blind evaluation with 163 experts across 203 engineering tasks, Hy4 preview scored 2.99/4.00, slightly ahead of GLM-5.

MyClaw's August 28, 2026 overview of Hy4 Preview restates the architecture as a 770B-total / 49B-active MoE with a 1M-token context window, a 78-layer backbone using 256 routed experts plus a shared expert, eight routed experts active per token, and a native multi-token prediction layer for speculative decoding. It fra The same post surfaces a vendor-acknowledged caveat that is genuinely useful for Hy4-preview adopters: Tencent warns the model can spend too long reasoning and over-verifying its work, and the article notes that independent reliability evidence is limited on release day. Because MyClaw's piece also doubles as a directo

BigGo Finance reported on August 28, 2026 that Tencent Hunyuan officially released and open-sourced its next-generation flagship large language model, Hy4 preview, with 770 billion total parameters, 49 billion active parameters, and a 1 million token context length. According to the report, Hy4 preview achieved a gener The article reports that Hy4 preview is priced at 6 yuan per million input tokens and 18 yuan per million output tokens, and is available on Tencent Cloud TokenHub and OpenRouter. It also notes that Hy4 preview represents a more-than-doubling of total parameters compared to the previous-generation Hy3's 295 billion, wi