Tencent's Hy team just released Hy3 — a 295B-parameter Mixture-of-Experts model that activates only 21B parameters per token, yet goes head-to-head with trillion-scale flagships. Apache 2.0. 256K context window. Built specifically for agentic use cases. And they're giving away the API free for two weeks.
Specs at a Glance
| Property | Value |
|---|---|
| Architecture | Mixture-of-Experts (MoE) |
| Total Parameters | 295B |
| Active Parameters | 21B |
| MTP Layer Parameters | 3.8B |
| Layers | 80 + 1 MTP layer |
| Experts | 192 experts, top-8 activated |
| Attention | 64 heads, GQA, 8 KV heads, dim 128 |
| Hidden Size | 4,096 |
| Context Length | 256K tokens |
| Vocabulary | 120,832 |
| License | Apache 2.0 |
| Deployment | vLLM, SGLang, BF16 |
What Makes Hy3 Different
1. Efficiency-to-Performance Ratio
The defining stat: 295B total but only 21B active. You get the knowledge breadth of a massive model with the inference cost of a mid-size one. Tencent claims it outperforms similarly-sized models and rivals open-source flagships with 2–5× more parameters. If true, this is the most compute-efficient frontier model available under Apache 2.0.
2. Agent-First Design
Hy3 was built after gathering feedback from 50+ production products. The post-training pipeline emphasized agentic capabilities — tool use, multi-step reasoning, following complex instructions. This isn't a chat model that happens to do agents; it was designed for agent workloads from the start.
3. 256K Context with MoE Routing
Long-context MoE models have historically struggled with attention routing degradation over long sequences. Hy3's 256K window with 192 experts suggests Tencent solved the routing stability problem at scale — each token activates only 8 of 192 experts, keeping the KV cache manageable even at extreme lengths.
4. Apache 2.0 — No Strings
This is the killer feature for commercial adoption. No custom license, no "research only" restrictions, no output censorship clauses. Apache 2.0 means you can fine-tune, distill, deploy commercially, and build products on top — same license terms as Llama. For sovereign AI deployments and enterprise use, this removes the biggest friction point.
Why This Matters for the Open-Source Landscape
Hy3 lands in a competitive field: DeepSeek-V3, Qwen3, Llama-4, and Mistral Large all occupy the same "MoE frontier, Apache/MIT licensed" space. But Hy3's active parameter count (21B) is notably lower than DeepSeek-V3's (37B) and Llama-4's (varies by variant), which means lower inference cost per token for comparable quality.
The 256K context window also puts it ahead of most Apache 2.0 competitors for agentic workloads that need to process large codebases, multi-document reasoning, or long conversation histories.
Deployment: Ready for Production
Hy3 ships with immediate support for:
- vLLM — production-grade serving with PagedAttention
- SGLang — structured generation for agent workflows
- Fine-tuning — full recipe provided
- Quantization — supported for edge deployment
- Free API — 2-week trial via OpenRouter and Tencent's own endpoint
Relevance to Delta V
For sovereign AI deployments and agent architectures, Hy3 is immediately interesting:
- Apache 2.0 + 21B active means you can run it locally on a single high-end GPU or modest cluster without licensing concerns.
- Agent-first post-training aligns with our work on multi-agent systems, tool-augmented pipelines, and autonomous workflows.
- 256K context unlocks whole-codebase analysis, long-document reasoning, and persistent agent memory — all critical for production agent systems.
We'll be testing Hy3 against our internal benchmarks for agent reliability, instruction following, and structured output quality.
Sources: Hugging Face: tencent/Hy3 · Tencent Hy Research · @TencentHunyuan on X · Announced July 6, 2026