Tencent introduces Hy4, a new open-weight large language model

Summary

Tencent released Hy4 Preview, an open-weight large language model (LLM) that processes text only, featuring 770 billion total parameters (49 billion active) and a 1 million token context window. The model is approximately 2.6 times larger than its predecessor, Hy3, released in July. Simon Willison examined the model's chat template and conducted a quick SVG generation test via OpenRouter.

Why it matters

The 1.56 TB open-weight model from a major Chinese lab narrows the gap with closed-source models and offers researchers and self-hosters a genuinely large MoE option. The increase in context window from 256K to 1M and active parameters from 21B to 49B is seen as a significant step forward in the capacity that open Chinese labs can provide. The 'high / no_think' binary reasoning switch is a design choice that warrants careful scrutiny, as it limits how much users can reduce computational power.

Highlights

• 770B total / 49B active MoE, 1M context, 1.56TB download on Hugging Face
• Only two reasoning modes are available: 'high' (default) and 'no_think' — no intermediate levels
• The reasoning trace in the Pelican-SVG test contains intentionally shortened, token-efficient English

Critical take

The article is essentially a link share based on a single anecdotal prompt test; it lacks benchmark data, visual processing capabilities, latency, or cost information. Consequently, the claim of 'large size increase' relies entirely on parameter counts rather than proven quality gains. The binary reasoning switch also raises the question of whether 'high' mode means 'always think deeply' without user control over depth, which could lead to resource waste on routine tasks.