Hy4 preview

IntuitionLabs.ai provides a technical analysis of Tencent Hy4 Preview, describing it as an open-weight large language model developed by the Tencent Hy Team and released on August 28, 2026. The model is a Mixture-of-Experts design with 770 billion total parameters, of which 49 billion are activated per token, and is re The report documents a roughly 2.6-fold parameter increase and 4-fold context-length increase over Hy3, which shipped on July 6, 2026 with 295 billion total and 21 billion active parameters. Access paths include self-hosted inference with vLLM or SGLang, Tencent Cloud TokenHub API, and third-party routing via OpenRoute

This analytical article by IntuitionLabs explicitly names Hy4 preview and documents the model as a 770-billion-parameter Mixture-of-Experts design with 49 billion parameters activated per token, released by the Tencent Hy Team on August 28, 2026 under Apache License 2.0. It confirms weights distribution as both a stand The piece adds the most rigorous caveat layer of the candidate set, flagging that Tencent's blind comparative evaluation was conducted internally and noting methodological limitations in the self-selected internal test suite. It frames Hy4 preview's benchmark standing as a snapshot rather than a settled third-party-aud

TechNode reports Tencent's open-source release of Hy4 preview, a 770-billion-parameter MoE model with 49 billion activated per token and a context window exceeding one million tokens, released under Apache 2.0. The model is the successor to Hy3, which had 295 billion total parameters, 21 billion active parameters, and The article confirms the 78-layer architecture with 256 routed experts plus one shared expert (eight routed plus shared activated per token), plus a separate multi-token prediction layer for speculative decoding. Developers can download both the full model and an FP8 version, and Tencent published serving instructions

This Wavect review explicitly names Hy4 preview and documents the August 28, 2026 release with 770 billion backbone parameters, 49 billion activated parameters per token, and a context window above one million tokens. It lists WorkBuddy, CodeBuddy, Tencent Cloud TokenHub, and OpenRouter as official access routes, with The article provides a short verified-answer fact table with buyer-implication columns covering capacity, context, and access, and notes Tencent's own label of this as an early version that can reason and verify longer than necessary. All statements are directly supported by the supplied page excerpt, and the review wa

A CodeToDeploy article on Medium reports that Tencent's Hy4 Preview was released on August 28, 2026, as a flagship Mixture-of-Experts model with 770 billion total parameters and 49 billion parameters activated per token. The model features a 1-million-token context window, native multi-token speculative decoding, gated The article states that Hy4 Preview launched with two weeks of free access through WorkBuddy and CodeBuddy, and is also available through Yuanbao, ima, and APIs such as Tencent Cloud TokenHub and OpenRouter. It positions Hy4 Preview as competitive with GLM 5.3, Kimi K3, and Qwen 3.8 on key benchmarks, framing it as a n

This Emergent news article explicitly names Hy4 preview and reports Tencent's August 28, 2026 release under the permissive Apache 2.0 open-source license. It confirms the preview designation as an early-access version allowing developers to test capabilities before full production release, and explains the Apache 2.0 t The article is thin on technical detail — it provides no architecture, benchmark, or pricing specifics beyond the license — and frames the release within Tencent's broader AI portfolio alongside the Hunyuan series. It explicitly names Hy4 preview throughout, satisfying the candidate title and ID match, but contributes

This Developers Digest write-up explicitly names Hy4 preview and confirms the August 28, 2026 open-source release at 770 billion total parameters with 49 billion active, Apache 2.0 licensing, and a one-million-token context window, with weights on Hugging Face, ModelScope, GitCode, and CNB. The article states the model The article offers a decision-intent framing for engineering teams, discussing DeepSeek-style sparse attention as the mechanism behind the 1M context and situating Hy4 preview as the largest open-weights release of August 2026. It references benchmark numbers (Terminal-Bench 85.4, DeepSWE 64.3) and OpenRouter pricing o