Tencent Ships Hyra-1.0, an Agent That Critiques and Rewrites Its Own Work

Tencent's AI lab has shipped Hyra-1.0, the first version of the Hunyuan Research Agent. Rather than returning a one-shot answer, Hyra loops over its own outputs, critiques them, and keeps revising until it hits a performance target. The system targets tasks where quality is measurable: machine learning experiments, scientific computation, and engineering optimization.

How the loop actually works

"Recursive self-improvement" has floated around AI theory for decades, but Hyra grounds it in something concrete. The agent does not rewrite its own weights. Instead, it runs a controlled iterative cycle: generate a solution, evaluate it against a verifiable objective, identify what failed, and produce a revised attempt. The model acts as both solver and critic, and the loop terminates on a metric rather than a word count or summary.

Most research agents are optimized for information retrieval and synthesis. Hyra is built for performance-driven tasks, which means the evaluation signal drives every iteration.

Three categories of use cases

Tencent organized Hyra's demos into three buckets:

• AI4AI: Using the agent to improve AI systems themselves, including hyperparameter search, architecture tuning, and automated ML pipeline optimization. The evaluation signal is a benchmark score or validation loss.
• AI4Science: Applying AI to accelerate scientific discovery across material science, drug design, and genomic analysis. Hyra targets tasks where a correct solution can be verified against physical laws or experimental data.
• AI4Fun: Lower-stakes creative and game-oriented tasks where the loop optimizes for user-defined quality signals, such as game-playing strategies or generative outputs scored by a reward model.

The model underneath

Hyra runs on Hy3, Tencent's flagship open-source model. Hy3 is a 295B-parameter Mixture-of-Experts (MoE) architecture with 21B active parameters and 3.8B MTP layer parameters. Because MoE models activate only a fraction of parameters per token, inference costs stay manageable despite the large total parameter count.

Tencent improved Hy3 primarily through better post-training rather than scaling raw parameter count. The team raised the quality and diversity of post-training data and expanded reinforcement learning, producing solid gains across reasoning, agentic, and long-context tasks that compete with larger flagship models. That agentic foundation is what makes the Hyra loop viable: the model needs to reliably call tools, follow structured formats, and reason over multi-step plans without drifting.

Where it performs well

The iterative loop adds the most value when:

• You can define a numeric objective such as loss, accuracy, score, or pass rate
• Each iteration is cheap enough to run many times
• The search space is too large for manual tuning
• You want a system that self-corrects without human intervention between rounds

In productivity scenarios covering coding, financial modeling, frontend design, and game development, Hy3 already functions as a reliable, cost-effective option. Hyra layers the iterative loop on top of those strengths.

Where it falls short

The recursive loop only works when the evaluation signal is reliable. Open-ended tasks like writing a persuasive essay or designing a user interface lack a ground-truth metric to optimize against. In those cases, Hyra reduces to a standard agentic model with extra overhead. Handling complex tasks, adapting to experimental scientific domains, and ensuring output reliability remain open challenges.

Compute cost is also a real constraint. Running multiple iterations of a 295B MoE model is expensive, so Hyra makes sense only when the cost of a wrong answer outweighs the cost of extra inference cycles.

Where Hyra sits in the broader field

Recursive self-improvement is moving from thought experiments to deployed systems. LLM agents now rewrite their own codebases, scientific discovery pipelines schedule continual fine-tuning, and robotics stacks patch controllers from streaming telemetry. Hyra joins systems like Google's AlphaEvolve and OpenAI's o-series reasoning models, all of which operate on some version of the generate-evaluate-improve loop.

Hyra's distinguishing angle is its explicit focus on research and engineering rather than general-purpose assistance. Systems like FunSearch and AlphaEvolve discover algorithms that feed back into AI development itself. Hyra applies the same paradigm to a broader set of scientific and engineering problems.

Access and pricing

Hyra-1.0 is available through the Tencent Hy research page, which hosts demos across all three use-case categories. The underlying Hy3 model is released under the Apache 2.0 license and is progressively appearing on third-party platforms including OpenRouter, Hermes, Kilo, Cline, and Cherry Studio.

For local inference, Tencent released 1-bit and 4-bit GGUF quantizations for llama.cpp, enabling runs on 128 GB single-GPU rigs without an 8x H20 cluster. On Tencent Cloud, Hy3 is priced at 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.25 yuan per million cache-hit input tokens. That cache-hit rate is particularly relevant for Hyra's iterative loop, where the same long context gets reused across multiple improvement rounds.

Hyra-1.0 is a first release, and Tencent is clearly investing in the recursive improvement paradigm as a long-term direction. For teams running ML experiments, scientific simulations, or engineering optimization pipelines with a clear evaluation harness already in place, it is worth testing now.