NextLUCA
← ← Back to blog

Tencent Hy4 Preview: A 770B-Parameter Open-Source Model With a Million-Token Context Window

Tencent has released Hy4 preview, an open-source large language model with 770 billion total parameters and a context window exceeding one million tokens. It arrives with a two-week free access window on Tencent's own platforms and benchmark results that edge out competing models in internal testing.

Tencent has released Hy4 preview, an open-source large language model with 770 billion total parameters and a context window exceeding one million tokens. It arrives with a two-week free access window on Tencent's own platforms and benchmark results that edge out competing models in internal testing.

Specs

  • Total parameters: 770B
  • Active parameters: 49B
  • Context window: Exceeding 1M tokens
  • Input price: USD 0.834 per million tokens
  • Output price: USD 2.501 per million tokens
  • Cache-hit price: USD 0.042 per million tokens
  • Throughput improvement: 31.8% over baseline

What's new

  • Hy4 preview is Tencent's open-source large language model with 770B total parameters and 49B active parameters.
  • The context window exceeds 1 million tokens.
  • In an internal blind evaluation with 163 experts across 203 engineering tasks, Hy4 preview scored 2.99 out of 4.00, ahead of Kimi K3 (2.94) and GLM-5.3 (2.92).
  • Training data was co-created with Tencent experts in software engineering, gaming, finance, and security.
  • The model can generate a playable game prototype from a single natural-language request.
  • Hy4 preview contributed to automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators, and inference optimizations raised end-to-end throughput by 31.8% over baseline.
  • It is accessible via Tencent Cloud TokenHub and OpenRouter APIs, and through WorkBuddy, CodeBuddy, Yuanbao, and ima.
  • The next batch of Hy4-series models is expected soon.

Who is it for

  • Developers building agentic coding tools who want access through WorkBuddy or CodeBuddy.
  • Teams evaluating model providers on cost, given the low cache-hit pricing of USD 0.042 per million tokens.
  • Engineers working on long-document or large-codebase tasks that benefit from a context window exceeding 1M tokens.
  • Users of Tencent's existing product suite (Yuanbao, ima) looking for an integrated upgrade path.
  • Organizations comparing Hy4 preview against GLM-5.3 or Kimi K3 on engineering-task benchmarks.

How to try it

Hy4 preview is free on WorkBuddy and CodeBuddy for two weeks after launch, and Hy3 remains free on both platforms until September 30. Beyond that window, it's accessible via Tencent Cloud TokenHub and OpenRouter APIs at USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million tokens on cache hits, or through the Yuanbao and ima apps.

Community take

  • Early OpenRouter usage numbers for Hy4 preview were reported as unusually high, reportedly surpassing GLM-5.3's usage over a longer period, with the low cache pricing seen as a draw.
  • One thread linked the model's automated self-optimization claims to broader forecasts about AI progress, while another worried that heavy token-density optimization could erode linguistic nuance.
  • A separate view held that LLMs are best suited to tedious optimization and coding work, leaving higher-level judgment to humans.
  • On the predecessor Hy3, feedback was mixed: one user flagged slow inference, while another reported strong results as a general-purpose agentic model.

Sources