Hyra-1.0 is Tencent Hunyuan's research agent, released July 21, 2026, built for scientists, engineers, and ML researchers who want an autonomous system that improves its own solutions rather than just answering prompts. Instead of a chat bot, it is a recursive self-improving scaffold that runs a continuous loop of proposing, testing, and scoring candidate solutions inside isolated sandboxes.
Core Features
- A lightweight Harness with a Context Agent that maintains an Experience Bank of past code, logs, and scores, then synthesizes inspirations for a task queue.
- Multiple Proposal Agents drain the queue, write each solution as a solve.sh entry point, and run it in a fresh sandbox where it receives an automated score.
- A bilevel loop for tasks without a built-in evaluator: Hyra first builds an evaluator from the description, then upgrades that evaluator using accumulated experience to fight reward hacking.
- An asynchronous producer-consumer pipeline that scales solution quality with time and compute instead of hitting a fixed benchmark ceiling.
- Open results published on GitHub, covering AI-for-AI, AI-for-Science, and AI-for-Fun tracks.
Use Cases
- AI-for-AI: it cut NanoChat Autoresearch validation BPB to 0.9015 and reached a 0.771 Mean SOL across 235 GPU kernels on SOL-ExecBench.
- AI-for-Science: it set new records on 29 of 55 math open problems and derived a sunspot recurrence formula with 0.77 out-of-sample R-squared on 1749-1932 data.
- Drug and hardware design: PARP1 inhibitor candidates scored higher drug-likeness than olaparib, and its qubit-routing algorithm beat classical SABRE by 44.4 percent on IBM Q20.
Pricing
Hyra-1.0 is a research release from Tencent Hunyuan with open results on GitHub; it is not sold as a commercial API tier. Access today is through the published harness and a recruitment questionnaire for collaborators rather than a pay-per-token plan.
Our Take
Best for research labs and engineering teams that want a transparent, sandbox-backed agent that improves its own outputs over time. The catch is that it is a scaffold, not a hosted product, so you supply the evaluators and compute. More at aifreetool.site/tool-category/ai-learning-tools/.
