Hy4 Preview (Tencent Cloud)

The AI Daily Digest for August 31, 2026 reports that Tencent released and open-sourced Hy4 preview on August 28 with API pricing set at $0.834 per million input tokens and $2.501 per million output tokens, with a two-week free window on WorkBuddy and CodeBuddy (both China and international builds), and Yuanbao and ima Tencent additionally claims that Hy4 preview participated in optimizing its own training methods, data strategy, evaluation frameworks, and low-level operators, then autonomously analyzed its inference bottlenecks to raise end-to-end throughput by 31.8 percent, framed as a recursive self-improvement loop. The Hunyuan c

Tencent released Hy4 preview as an open-weight mixture-of-experts model on 28 August 2026, publishing the checkpoint, an FP8 quantised variant, a fine-tuning pipeline and serving recipes on Hugging Face under an Apache 2.0 license. The headline architecture is 770 billion total parameters with 49 billion activated per The progressiverobot breakdown frames Hy4 preview as Tencent's new flagship MoE and traces every released figure back to its references, including benchmark numbers and the 1M-token context window. The article covers what was shipped (weights, FP8 variant, fine-tuning pipeline, serving recipes), how the architecture wo

Tencent Cloud's techpedia article dated 28 August 2026 documents Hy4 Preview's specs as a 1M-token context window, 960k maximum input, and 64k maximum output, and positions the model for coding scenarios including repository-level analysis. The piece argues that those specifications make repository-level analysis worth The guidance covers defining the analysis unit before collecting code by anchoring a repository manifest to a commit SHA, constructing evidence graphs outside the model, using a layered index to assemble direct context, keeping claims traceable and falsifiable, and letting builds and tests challenge the analysis with c

Tencent Cloud's techpedia article dated 28 August 2026 uses Hy4 Preview to describe a research-agent scenario that coordinates multiple Codex-style sessions and iterates from their results, explicitly labelling this as product-direction evidence rather than proof that Hy4 contains an internal experiment scheduler. A cr The workflow guidance walks through encoding a study as a versioned protocol tied to a hypothesis, baseline, treatment, fixed workload and primary endpoint, isolating concurrent sessions by research purpose, making each run independently inspectable, preserving failures and contrary evidence, and choosing the next expe

Tencent released Hy4 preview on August 28, 2026 as a 770-billion-total-parameter mixture-of-experts model with 49 billion parameters active per token and a context window of one million tokens. According to the progressiverobot.com breakdown, the release ships with weights on Hugging Face under an Apache 2.0 license, a The same breakdown details the backbone as 78 layers with a hidden size of 6,144 and a vocabulary of 120,832 tokens, and frames Hy4 preview as Tencent's flagship open-weight entry into a crowded month that also included a 125B multimodal MoE preview from Alibaba and the unmasking of Ox Alpha as a Z.ai model. Tencent po