technode.com web signal

Tencent open-sources Hy4-preview, a 770B MoE with 1M-token context

By Published Updated
5 sources tracking this story
China AI Open Source Generative AI Inference ai-business

TL;DR

  • SWE-bench Pro 65.7 and GPQA Diamond 92.3 at $0.834 per million input tokens place Hy4 preview at frontier performance with pricing below most Western counterparts.
  • The 49B active parameter design on a 770B total keeps per-token inference compute closer to a mid-size dense model, compressing the cost floor further.
  • Hy3 saw 68-fold weekly usage growth after converting from preview to formal release, giving Tencent a replicable commercial ramp template for Hy4.

Tencent's Hunyuan team released Hy4-preview on August 28, a 770-billion-parameter mixture-of-experts model with 49 billion activated per token and a context window exceeding one million tokens, TechNode reported. The weights ship under Apache 2.0.

The architecture: 78 layers, the first a standard dense feed-forward network, the remaining 77 running MoE with 256 routed experts and one shared expert.

API pricing on Tencent Cloud TokenHub is $0.834 per million input tokens and $2.501 per million output tokens, with a two-week free trial on WorkBuddy and CodeBuddy and additional API access via OpenRouter.

Tencent says the model helped tune its own stack. Per the release, "the model helped optimize parts of its own training and inference systems, increasing end-to-end throughput by 31.8% against a baseline." On an internal rubric graded by 163 experts across 203 engineering tasks, Hy4-preview scored 2.99 out of 4, ahead of GLM 5.3 at 2.92 and Kimi K3 at 2.94.

Those are self-reported numbers on a self-designed rubric; the TechNode write-up carries no GPQA, SWE-bench, or other public benchmark scores. Hy4-preview joins a steady run of Chinese open-weight releases we've tracked this quarter, one of more than 340 China AI stories in our tracker over the last 90 days.

What others are reporting

Coverage cluster as of 24h after publish

  1. Tencent Read →

    First-party release with official benchmark table, API pricing tiers, and two-week free access window on WorkBuddy and CodeBuddy.

    Ranked among the top tier of open-source models, Hy4 preview delivers outstanding performance across coding, office work, and scientific research.
  2. Crypto Briefing Read →

    Surfaces the 68x Hy3 usage spike post-launch and Tencent's parallel strategy of running proprietary models alongside open-source DeepSeek internally.

    Weekly usage of Hy3 rose more than 68-fold after it transitioned from preview to formal release.
  3. MindStudio Read →

    Technical breakdown of Gated DeepSeek Sparse Attention with IndexCache, 4-stream Hyper-Connections, and the native MTP layer built for speculative decoding.

    Hy4 preview is an open-weight Mixture-of-Experts language model from the Tencent Hy (Hunyuan) Team, released under Apache 2.0.
  4. Enera Labs Read →

    Enterprise deployment framing: cautions against production coding agents pre-stabilization, recommends workflow testing during free window, flags GLM-5.3 as current coding leader.

Originally reported by technode.com

Read the original article →

Original headline: Tencent Open-Sources Hy4-Preview, a 770B MoE With 49B Active and 1M Context Under Apache 2.0