Hy3

Architectural Overview & Capability Profile

Hy3 is developed by Tencent as an advanced foundation model engineered for production enterprise applications, reasoning workflows, and autonomous problem-solving. On the comprehensive RankLLMs composite evaluation index, Hy3 registers an overall benchmark rating of 42.9 / 100.

Operating under a Open Weights license structure, Hy3 is distributed as an open-weights architecture capable of local deployment, private VPC inference, and custom fine-tuning.

Benchmark Performance & Real-World Utility

In standardized laboratory and real-world evaluation suites, Hy3 demonstrates calibrated proficiency across key performance vectors:

• SWE-bench Verified Software Engineering (32.9%): Measures multi-file GitHub issue resolution. The model demonstrates practical capability in localizing source code bugs, synthesizing functional patches, and passing rigorous integration test matrices. For evaluation methodology, consult our SWE-bench Pro guide.

• GPQA Diamond Scientific Reasoning (42%): Assesses graduate-level domain problem solving across physics, chemistry, and biology, minimizing hallucinations in multi-step deductive arguments. See our Humanity’s Last Exam benchmark analysis.

• MATH-500 Competition Mathematics (42%): Evaluates formal mathematical problem solving and symbolic derivation, providing foundational logic for algorithmic computing.

• Terminal-Bench Execution (70.8%): Reflects proficiency in interactive command-line environments, shell tool use, and automated DevOps scripting.

Inference Latency & Cost Architecture

• Token Context Window: 256K tokens, allowing large documents, multi-file codebases, and long conversation memories to be processed in a single context window.
• Inference Speed: Reference throughput clocks at approximately 50 tokens/second, enabling responsive streaming in developer tools and interactive interfaces.
• Commercial Pricing: Zero per-token inference fee when self-hosted on owned GPU infrastructure (e.g. vLLM or Ollama), with marginal costs limited strictly to hardware compute.

Strengths, Limitations & Ideal Workflows

Primary Strengths

• Open source foundation weights under permissive licensing
• 70.8 on Terminal Bench 2.1 and 56.2 on Toolathlon
• Self-hostable local enterprise architecture

Known Considerations

• Lower coding pass rates on modern benchmarks (32.9 SWE Atlas / 28.0 DeepSWE) compared to next-generation Hy4

Recommended Use Cases

Legacy self-hosted deployments, localized enterprise inference, and custom fine-tuning.

Related Comparisons & Leaderboard Context

• Compare with alternatives in the Model Comparison Engine
• Compare global rankings on the AI Model Leaderboard
• View complete platform index in our HTML Sitemap & Directory
• Explore top coding models in our Best LLM for Coding Guide