Hy3 — AI Model Review & Pricing
Apache 2.0 licensed Mixture-of-Experts agentic model with 295B total parameters (21B active), ~256K context window, and OpenRouter pricing at $0.14/M input tokens. Excels in agentic search (BrowseComp 84.2) and tool orchestration (MCP-Atlas 79.1), though trails GLM-5.2 on coding benchmarks. Text-only modality.
Specifications
| Attribute | Value |
|---|---|
| Lab | Tencent |
| Tags | Reasoning Agentic Coding Open Weight |
| Release Date | 2026-07 |
| Context Window | 262,144 tokens |
| Input Price / 1M | $0.14 |
| Output Price / 1M | $0.58 |
| Input Modalities | Text |
| Output Modalities | Text |
Strengths
- 84.2 on BrowseComp for agentic search
- 79.1 on MCP-Atlas for tool orchestration
- Apache 2.0 license for open-weight use
- ~256K context window for long documents
Weaknesses
- 78.0 on SWE-Bench Verified trails GLM-5.2 (84.2)
- Text-only input without vision or audio support
- High GPU requirements (8 GPUs for self-hosting)
- Independent benchmark verification pending
Best For
- Agentic search and information retrieval
- Tool orchestration and multi-step workflows
- Long-context document analysis
- Cost-sensitive production deployments
In Depth: Hy3
Benchmark Performance
Hy3 agentic model scores LMRank 8.8/10 (#24), anchored against GPT-5.5 Pro at 9.5/10 (#5). Excels in structured tasks where tool use and long-context reasoning matter.
Tencent reports Hy3 achieves 84.2 on BrowseComp for agentic search and 79.1 on MCP-Atlas for tool orchestration, outperforming GLM-5.2 in these domains. However, on SWE-Bench Verified it scores 78.0 versus GLM-5.2's 84.2. These figures are vendor-reported and await independent verification.
Pricing & Value
Hy3 costs $0.14 per million input tokens and $0.58 per million output tokens on OpenRouter, with a two-week free API period. Tencent Cloud TokenHub pricing is similar: 1 yuan ($0.14) input and 4 yuan ($0.55) output.
At $0.14/M input, Hy3 is over 71x cheaper than GPT-5.5 Pro ($30.00/M input). Even output at $0.58/M is 310x less than GPT-5.5 Pro's $180.00/M output. For price-per-LMRank-point, Hy3 costs $0.016 per input point, versus GPT-5.5 Pro's $3.16—making Hy3 a high-value option for budget-constrained agentic workloads.
Who Should Use This
Hy3 is ideal for developers building agentic workflows, teams needing long-context processing on a budget, and organizations requiring an open-weight model with permissive licensing.
- Agentic workflow builders: excels in tool orchestration and search, but coding benchmarks are not top-tier
- Cost-conscious teams: 71x cheaper input than GPT-5.5 Pro, though text-only modality limits versatility
- Open-weight adopters: Apache 2.0 license enables self-hosting, but requires 8 GPUs
- Not for: vision or multimodal tasks—Hy3 is text-only
Release & Version History
Hy3 by Tencent launched on July 6, 2026 as the final release after an April 2026 preview. It is the successor to Hy2 and the fully trained version of the Hy3 Preview.
Tencent released Hy3 Preview in April 2026, offering an early look at the Mixture-of-Experts architecture with 295B total parameters (21B active). The final Hy3 model arrived on July 6, 2026, incorporating improvements in hallucination reduction and multi-turn consistency. A two-week free API period followed launch, and the open-weight release under Apache 2.0 enables community self-hosting.