Test Tencent Hy3 before choosing your next open agent model
Tencent Hy3 is now public on Hugging Face and OpenRouter with Apache-2.0 weights, a 295B MoE design, 21B active parameters, and a 256K-class context window.
Tencent Hy3 is confirmed as a public open model release on Hugging Face and is listed on OpenRouter for API access. Confidence level: confirmed for public model availability, Apache-2.0 licensing, 295B total parameters, 21B active parameters, and 256K-class context. Benchmark and production-reliability claims are still vendor-reported or early third-party summaries, so teams should reproduce them before replacing GLM-5.2, Qwen, DeepSeek, or closed coding models.
What changed
Tencent's Hy team has published Hy3 as a 295B-parameter Mixture-of-Experts model with 21B active parameters and an Apache-2.0 license. The Hugging Face model page lists the model as public, ungated, text-generation focused, and tagged with license:apache-2.0.
OpenRouter also lists tencent/hy3 as a public model route with reasoning support, a 262,144-token context length, and provider pricing. That makes Hy3 more than a paper or benchmark post: developers can inspect weights, test hosted API routing, or plan self-hosting with vLLM or SGLang.
| Model path | Best fit | Access | Cost/status | Caveat |
|---|---|---|---|---|
| Hy3 on Hugging Face | Weight inspection, self-hosting, reproducibility | Public model page | Apache-2.0 weights | Large 295B total model still needs serious serving hardware |
| Hy3 on OpenRouter | Fast hosted testing and routing experiments | Public OpenRouter listing | Provider pricing, route-dependent | Provider limits and context support can differ from self-hosted behavior |
| Hy3-FP8 | Lower-memory serving tests | Separate quantized checkpoint | Smaller serving footprint | Quantization quality needs task-level validation |
| GLM-5.2 | Long-context coding-agent baseline | Z.ai and model distribution pages | Open model ecosystem | Still appears stronger on several coding benchmarks |
Why this is early
Hy3 surfaced quickly through model-distribution and tracker lanes: Hugging Face, OpenRouter, AI media, and community model-watch channels picked it up before slower enterprise writeups. The public artifacts are real, but evaluation is still early because most numbers come from Tencent's own model card or summaries of Tencent's appendix.
The practical claim is clear: Hy3 may be cheaper to run than larger open MoE systems while staying useful for agentic workflows, long-context work, document processing, financial analysis, frontend work, and coding support. Those are exactly the claims teams should test against their own repos and workflows.
Key takeaways
- Hy3 is public on Hugging Face with Apache-2.0 licensing and no visible gating on the model page.
- Tencent describes Hy3 as a 295B MoE with 21B active parameters and 3.8B MTP layer parameters.
- OpenRouter lists Hy3 with a 262K-token-class context window and reasoning-effort support.
- Independent technical coverage frames Hy3 as smaller and more commercially permissive than GLM-5.2, while noting GLM-5.2 still leads on coding benchmarks.
- Treat benchmark tables as starting points, not deployment evidence.
Availability and access
Developers can inspect Hy3 on Hugging Face now. The model page points to Transformers, vLLM, SGLang, Docker Model Runner, and quantization paths. It also lists Apache-2.0 licensing and public model metadata.
OpenRouter lists Tencent Hy3 as an API-routed model with text input/output, reasoning support, and provider-level pricing. Use those listings for quick trials, but verify actual context limits, max output, tool support, latency, data policy, and pricing on the provider route you select.
Practical LinkLoot angle
Hy3 is worth testing if you want an open-weight model for long agent sessions, document-heavy analysis, or lower-cost hosted experiments. Do not start with a leaderboard. Start with one workflow that currently costs too much or fails under context pressure.
Good first tests:
- A multi-file repo issue that requires planning, editing, and verifying.
- A long document or contract where fabricated clauses would be expensive.
- A finance or operations analysis that must flag missing data instead of guessing.
- A frontend task where layout fidelity and follow-up edits matter.
- A tool-calling flow where schema stability matters more than a single benchmark score.
For broader evaluation patterns, use LinkLoot's AI agent tools guide: /guides/ai-agent-tools.
What to verify before you act
Check the Hugging Face license file, model card, files, and update history before treating Hy3 as approved for commercial use. Also verify whether your hosting path uses the full model, FP8 weights, a provider-specific quantization, or a routed endpoint with hidden limits.
Run your own benchmark set before switching a coding agent. The model card and independent summaries both show a useful pattern: Hy3 looks strong for open, agentic, and long-context work, while GLM-5.2 may remain ahead on several coding tasks. Your codebase, test harness, and review burden matter more than averaged scores.
Review data controls before sending proprietary repositories or documents through a hosted route. OpenRouter provides provider-level policy details, but those can vary by endpoint. For self-hosting, check GPU memory, throughput, vLLM/SGLang recipe maturity, and fallback behavior under long context.
Source check
Confirmed by: Hugging Face confirms the public tencent/Hy3 model page, Apache-2.0 tag, model metadata, and deployment guidance. OpenRouter confirms public model availability, routing metadata, context-window class, reasoning support, and provider-level pricing details.
Early signal / context: MarkTechPost summarizes Tencent's model-card claims, benchmark comparisons, Apache-2.0 licensing, FP8 checkpoint, and GLM-5.2 comparison. The Hugging Face page includes executable setup examples, so LinkLoot treated those snippets as untrusted source text and verified the factual claims against OpenRouter and independent coverage before publishing.
Hy3 is Tencent's open Mixture-of-Experts language model with 295B total parameters and 21B active parameters per token.