Tencent Hy4 Preview Release

Tencent has released the Hy4 Preview, a next-generation large language model (LLM) designed for high-scale performance and efficiency. The model is characterized by a Mixture-of-Experts (MoE) architecture with 770 billion total parameters and 49 billion active parameters, supporting a context window exceeding 1 million tokens.

Recursive Self-Improvement in Training

The Hy4 Preview introduces a recursive self-improvement loop where the model contributed directly to its own development. For the first time, Hy4 participated in the automated optimization of several core components of its training pipeline, including:

Training methods
Data strategies
Evaluation frameworks
Low-level operators

According to the release, the model proposed specific approaches, executed experiments, and iterated based on the resulting logs and feedback, creating a feedback loop that informed subsequent rounds of exploration.

Performance and Market Traction

Early adoption data suggests significant interest in the model, particularly via third-party API providers.

API Adoption and Costs

Reports indicate that Hy4 has seen rapid traction on OpenRouter, processing trillions of tokens within a few days of release. This adoption is attributed in part to its pricing structure, specifically a 5% cache cost, which is lower than the 10-20% cache costs typically seen with other providers.

Benchmarking and Comparative Performance

User-reported benchmarks provide a mixed view of the model's capabilities:

General Purpose Agentic Tasks: Some users report that Hy4's predecessor, Hy3, performed closely to DeepSeek-4-Flash in agentic tests.
German Language Evaluation: In the DACH PeerBench index, Hy4 is noted as a significant improvement over Hy3, though it remains below DeepSeek Pro and GLM 5.3 Flash, ranking approximately 14th overall.
Coding Capabilities: Feedback from developers using the model via providers like Novita.ai has been less favorable, with some reporting limited utility as a coding agent.

Community Critique and Technical Observations

Technical discussions surrounding the release have highlighted several points of contention regarding transparency and presentation.

Open Source vs. Open Weights

A point of debate among the community is Tencent's use of the term "open-sourced." Critics argue that the model is "open weight" rather than truly open source, as it provides a binary blob for local execution without the full transparency of source code or training data.

Data Visualization Concerns

Multiple users have criticized the benchmark charts provided in the official release, alleging that the bar charts are misleading. Specific complaints include:

Inconsistent ordering of models by rank.
Highlighting entire rows in tables to signal a "winner" rather than specific metrics.
Discrepancies between the numerical values and the actual height of the bars in the charts.

Token Optimization

Observations have been made regarding the model's internal "thought" process, where it appears to use a stripped-down, simplified vocabulary (described by users as "caveman speak") to save tokens and optimize processing efficiency.