Tencent Hy4 Preview Release: 770B Model with 1M Token Context and Open-Weight Code

Tencent Hy4 Preview – What It Is and Why It Matters

Tencent announced the Hy4 preview, a next‑generation large language model (LLM) with 770 billion total parameters, of which 49 billion are active during inference, and a context window exceeding 1 million tokens. The company also released the model’s open‑weight code, allowing anyone to run the model locally rather than through a proprietary API. This release signals that Chinese AI labs are now competing at the frontier of model scale and openness.

Core Technical Specs

Parameter count: 770 B total, 49 B active (the portion used for inference).
Context length: >1 M tokens, far surpassing the typical 8‑32 K token windows of most contemporary LLMs.
Open‑weight release: Tencent published the model’s weight files and the associated inference code, enabling independent deployment on custom hardware.
Training innovations: The Hy4 preview was the first Tencent model to participate in its own training loop—proposing data‑selection strategies, running experiments, and feeding results back into subsequent optimization cycles, creating an early‑stage recursive self‑improvement (RSI) loop.

Early Adoption and Performance Signals

OpenRouter traffic: Within a few days of launch, the Hy4 preview handled trillions of tokens on OpenRouter, outpacing GLM‑5.3’s weekly token volume.
Cost efficiency: Providers report a 5 % cache cost for Hy4, compared with the 10‑20 % cache rates typical for other large models, making it financially attractive for high‑throughput applications.
Benchmark impressions: Users on Hacker News noted that Hy4’s performance is comparable to DeepSeek‑4‑Flash on general‑purpose tasks, though some observed slower inference speed relative to smaller models.

Community Reactions and Critiques

"Hy4 apparently has ludicrous traction on OpenRouter already, with trillions of tokens processed in a couple days: more than GLM‑5.3 in a week. That said, it's relatively cheap with a 5 % cache cost when everyone is still doing 10 %/20 % cache costs, so Hy4 may be more compelling." – minimaxir

"Hy4 preview also contributed to its own development process, participating for the first time in the automated optimization of training methods... This established an early‑stage recursive self‑improvement loop." – codethief

"I experimented with Hy3 for a project and was surprised with how good it was. It was only beaten by DeepSeek‑4‑Flash in our tests. It felt so close to DeepSeek behaviour I kept thinking it must have been forked from it." – jorl17

"Open weight is not the same as open source. The current 'open weight' models are just opaque binary blobs you can run on your own computer instead of through a web API." – petcat

Key takeaways from the discussion:

Performance parity: Early testers find Hy4 competitive with top‑tier models but note latency concerns, especially on public inference endpoints.
Open‑weight semantics: Some community members argue that releasing weight files does not equate to true open‑source software, as the code remains a compiled binary without full transparency.
Benchmark presentation: Several commenters criticize the blog’s visualizations, claiming bar charts are misleading and that highlighted rows may overstate Hy4’s advantage.
Geopolitical context: The release is viewed as a milestone in China’s AI race, with some observers noting that Chinese labs have already caught up to or surpassed Western counterparts.

Practical Implications for Developers

Large context windows enable new use‑cases such as processing entire books, long codebases, or massive logs in a single prompt, reducing the need for chunking strategies.
Cost‑effective caching makes Hy4 attractive for high‑throughput services like chat assistants, document summarization, or batch inference pipelines.
Self‑optimizing training loop suggests future models may become increasingly autonomous in improving their own data pipelines, potentially accelerating iteration cycles.
Deployment considerations: While weights are available, the model’s size demands multi‑GPU or specialized hardware; latency remains a challenge on public endpoints.

Outlook

Tencent’s Hy4 preview demonstrates that Chinese AI firms are now delivering frontier‑scale models with open‑weight releases, challenging the dominance of U.S. labs. The combination of massive parameter counts, unprecedented context length, and a self‑optimizing training pipeline points to a new generation of LLMs that can handle more complex, long‑form tasks while reducing operational costs. However, the community’s mixed reactions—particularly around true openness, benchmark transparency, and inference speed—highlight the ongoing trade‑offs that developers must weigh when adopting cutting‑edge models.