tencent · hy_v3
Hy3
Hy3 by tencent: a 298.8B-parameter mixture-of-experts open-weight model under the apache-2.0 license.
981,065 downloads · 995 likes · Model card · synced 2026-11-19
Specifications
- Parameters
- 298.8B (12.5B active)
- Architecture
- moe · hy_v3
- Layers
- 80
- Hidden size
- 4,096
- KV heads · head dim
- 8 · 128
- Vocabulary
- 120,832
- Native dtype
- —
- Context window
- 262,144 tokens
- Released
- 2026-07-02
- Last modified on Hub
- 2026-07-20
Features & licensing
- License
- apache-2.0
- Commercial use
- Yes
- Modalities
- text
- Tool / function calling
- No
- Reasoning mode
- No
- Languages
- EN
- Quantised variants on Hub
- FP8, GGUF, bitsandbytes
- Pipeline
- text-generation
Task fit
Editorial scores (0–100) used by the recommendation engine.
- Chat70
- RAG / Q&A70
- Code70
- Summarisation70
- Extraction70
- Agents70
Hardware to run Hy3
Weights need about 657 GB at FP16, 345 GB at INT8 and 197 GB at INT4. The KV cache adds roughly 328 MB per 1,000 tokens per request. For a reference workload of 5 requests per second with 1,500 input and 300 output tokens, total GPU memory is around 341.4 GB, which fits on 5 × NVIDIA A100 (80 GB).
| Configuration | VRAM | Utilisation | Est. first token | Cloud / month |
|---|---|---|---|---|
| 8 × NVIDIA L40S (48 GB) | 384 GB | 89 % | ~0.2 s | $11,096 |
| 5 × NVIDIA A100 (80 GB)recommended | 400 GB | 85 % | ~0.1 s | $11,680 |
| 5 × NVIDIA H100 (80 GB) | 400 GB | 85 % | ~0.0 s | $16,425 |
Estimates only. Run an assessment for your own traffic.
Auto-imported from Hugging Face. Architecture, license and feature data are synced from the model repository; the task-fit scores below are provisional defaults until a curator reviews them.