Within days of its release, Tencent’s Hy4 preview had reached #10 on OpenRouter’s model rankings. As of August 30, 2026, it had processed 3.07 trillion tokens through the platform. That is a remarkable adoption signal for a newly released model, although it needs to be read correctly: OpenRouter ranks models by tokens processed on its own platform, not by quality, number of users or share of the entire LLM market.
A Large Model Built to Run Efficiently
The model itself deserves attention. Released by Tencent on August 28 under the Apache 2.0 licence, Hy4 preview is a mixture-of-experts model with 770 billion total parameters, of which 49 billion are active for each token. It supports a one-million-token context window.
Its architecture brings together several ideas aimed at making a very large model practical to run. Tencent’s model card describes Gated DeepSeek Sparse Attention, IndexCache for reusing sparse-attention indexes across layers, and a native multi-token-prediction layer for speculative decoding. In plain language, these mechanisms try to avoid doing unnecessary work and to generate output faster.
But those specifications are not the most interesting part of the release.
Model-Assisted Model Engineering
Tencent says Hy4 participated in its own development: helping optimize training methods, data strategies, evaluation frameworks and low-level inference operators. According to the company, the model proposed approaches, ran experiments and used the resulting code, logs and feedback in subsequent rounds. It also analysed bottlenecks in its inference system and worked on areas such as operator fusion and communication optimization.
Tencent calls this an “early-stage recursive self-improvement loop.” I would use the term more cautiously. Tencent’s published account does not show Hy4 independently deciding what it should become, redesigning itself without supervision and releasing a better successor. The objectives, infrastructure and validation process remained part of a human-designed engineering system.
What Tencent describes is better understood as model-assisted model engineering. That may sound less dramatic, but it is consequential enough.
The engineering of an LLM extends far beyond designing the neural network. Teams must curate data, construct evaluations, run experiments, inspect failures, optimize GPU kernels and tune the serving stack. If a capable model can participate across this entire pipeline, the productivity gain is no longer confined to the people using the finished product. It enters the process that creates and operates the product.
Tencent reports that model-assisted inference optimization improved end-to-end throughput by 31.8% over its baseline, with gains across different context lengths and concurrency levels. This is a publisher-reported result, not an independently validated one. We also do not yet have enough detail to determine how much of the improvement came from the model, the human engineers or the workflow joining them.
The Productivity Multiplier Moves Upstream
Still, the direction matters more than this particular percentage.
Hy4 is not an isolated example. In 2025, Google DeepMind reported that AlphaEvolve, a Gemini-powered coding agent, improved a kernel in Gemini’s architecture by 23%, reducing overall training time by 1%.
Software has always been used to build better software. LLMs add something different: they can work with intent, source code, experimental results and natural-language feedback in the same loop. That potentially compresses the distance between forming a hypothesis, implementing it, testing it and learning from the result.
The immediate outcome is faster ML engineering. The second-order effect is more important: the productivity multiplier may affect the rate of AI progress itself, not merely end-user productivity. Hy4 is not proof of autonomous recursive self-improvement. It may, however, be an early and concrete example of the engineering loop beginning to tighten.
—
Follow me on X/anil for daily updates.