Hy4 preview

IntuitionLabs' 5 September 2026 analysis treats Hy4 preview as a Mixture-of-Experts design with 770B total parameters and 49B activated per token, released by the Tencent Hy Team on 28 August 2026 under Apache License 2.0, with weights distributed as both a full-precision Hy4-preview checkpoint and an FP8-quantized Hy4 IntuitionLabs explicitly cautions that Tencent's blind comparative evaluation was conducted internally and that the detailed results carry methodological limitations spelled out in a dedicated "Evaluation Methodology and Limitations" section, meaning Hy4 preview's benchmark standing should be read as a snapshot of a se

Wavect's review dated 1–2 September 2026 describes Hy4 preview as worth a controlled coding-agent pilot for teams constrained by long context, multi-step tool use, or repeated repository retrieval, while explicitly not recommending it as a safe default for every production workflow. The model is characterized as enormo The pilot work tested Hy4 preview inside WorkBuddy, and the review pairs its hands-on findings with the headline specs (770B backbone parameters, 49B active per token, context window above one million tokens) and the official access routes Tencent listed (WorkBuddy, CodeBuddy, Tencent Cloud TokenHub, OpenRouter), inclu

Developers Digest's 31 August 2026 write-up confirms Tencent's Hy4 preview as an Apache 2.0 open-weights MoE with 770B total parameters, 49B active per token, and a 1M-token text-only context window, posted with weights on Hugging Face, ModelScope, GitCode, and CNB. It verifies Terminal-Bench at 85.4 and DeepSWE at 64. The article scales Hy4 preview against Hy3 (295B total, 21B active, 256K context), giving 2.6× growth on both total parameters and context length, and treats the August 2026 release as the largest open-weights launch of the month. It cross-references the Hy4-preview model card, the OpenRouter listing, the Tencent Hy re

AI Profit Boardroom's 31 August 2026 coverage frames Hy4 preview as a next-generation Hunyuan model with 770B total parameters, 49B active per token, and a context window beyond one million tokens, open-sourced on 28 August 2026 and made available free for a two-week period inside WorkBuddy and CodeBuddy. It positions The piece credits Tencent's announcement and launch reporting for the productivity positioning (coding, office tasks, data analysis, game development, scientific research) and the benchmark framing as a generational jump on agentic coding. It is written as a buyer-facing guide that breaks down what shipped, what the be

Progressiverobot's 28 August 2026 coverage frames Hy4 preview as Tencent's flagship Mixture-of-Experts release with 770B total parameters, 49B activated per token, and a 1M-token context window, shipped as an open-weight checkpoint under Apache 2.0 with an FP8-quantized variant, fine-tuning pipeline, and serving recipe The piece walks through what the 1M-token window is actually for (large repositories, document collections, extensive logs, and long-running agent sessions) versus what it does not guarantee (perfect recall), and previews the benchmarks, hosting costs, and caveats the rest of the article expands on. It positions the re

Tencent's official announcement page confirms the open-source release of Hy4 preview on 28 August 2026 as a next-generation large language model with 770B total parameters, 49B active parameters, and a context window exceeding 1M tokens, released under an open-source license and positioned among the top tier of open-so The announcement cites a Tencent-conducted blind evaluation with 163 experts across 203 engineering tasks in which Hy4 preview scored an average of 2.99/4.00, narrowly ahead of GLM-5.3 at 2.92/4.00 and Kimi K3 at 2.94/4.00, attributing gains to high-quality training data co-created with Tencent experts in software engi

MyClaw's 28 August 2026 write-up summarizes Hy4 preview as the Tencent Hy team's new flagship Mixture-of-Experts language model, with 770B total parameters, 49B activated per token, a 1M-token context window, 78 layers, 256 routed experts plus a shared expert (top-8 routing), and a native multi-token prediction layer f The article stresses that the 1M-token window enables large repositories, document collections, extensive logs, or long-running agent sessions but does not guarantee perfect recall and increases input cost, and recommends empirically testing whether the model finds the right evidence, follows buried instructions, and m

The Tencent Hy Team's Hy4 preview model card on Hugging Face confirms the model as a Mixture-of-Experts (MoE) flagship with 770B total parameters and 49B activated per token, organized as 78 layers (1 dense FFN plus 77 MoE layers) with 256 routed experts (top-8) and 1 shared expert per layer, plus a native 10B MTP laye The model card positions Hy4 preview at the open-source frontier and emphasizes a substantially larger post-training run as the driver of capability gains, targeting long-context development, office productivity, and scientific workloads. Architectural choices mirror and extend recent advances from DeepSeek and GLM (sp