Hunyuan A13B Instruct
tencent · released 2025-07-08 · other license
Hunyuan-A13B-Instruct is Tencent's June 2025 MoE (80B total / 13B active) with dual-mode reasoning and 256K context, under the Tencent Hunyuan Community License: GPQA 71.2, MATH 94.3, AIME 76.8, and a leading 78.3 BFCL for agentic tool use (vendor).
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 80.39B total · MoE, — active |
| Architecture | hunyuan_v1_moe |
| Context window | 131K tokens |
| Knowledge cutoff | 2025-03-31 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 71.2% |
|---|---|
| SWE-bench Verified | — |
| AIME | 76.8% |
| MMLU-Pro | — |
| BFCL v3 (tool use) | 78.3% |
| Composite score | 7.65 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 48.1 GB | 46.6 GB | — | 131K |
Strengths & weaknesses
Strengths: Very efficient MoE (80B total / 13B active) with competitive reasoning, math and agentic scores; Hybrid fast/slow thinking, 256K context; Strong agentic tool use (BFCL 78.3)
Weaknesses: Tencent Community license with use restrictions (not Apache/MIT); AIME 2025 (76.8) trails leading reasoners