Hy4-preview
Tencent·
LLMsopen-weightFrontier
Overview
Preview of Tencent's next-generation Hunyuan flagship: a 770B-parameter MoE activating 49B per token, with a 1M-token context window. Open-weight under Apache 2.0 with an FP8 checkpoint alongside BF16, tuned for coding agents, complex tool-use workflows and productivity tasks, and trained on domain data from Tencent's software, gaming and finance teams.
Capabilities and innovations
770B total / 49B active MoE1M token context windowCoding-agent and tool-use focusApache 2.0 licenseTencent-internal software, gaming and finance domain dataFP8 checkpoint shipped alongside BF16
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 92.3% | vendor GPQA Diamond 92.3, vendor model card. |
| SWE-bench Pro | 65.7% | vendor SWE-bench Pro 65.7, vendor model card. |
| Terminal-Bench 2.x | 55.06% | independent Terminal-Bench 2.1 55.06%, measured by Vals AI with the Terminus 2 harness (accessed 2026-09-23). Vals AI is the reference harness for this field so values compare across vendors; vendor and other third-party figures stay in the notes below. No Terminal-Bench 2.x value was recorded before. |
- Humanity’s Last Exam (no tools): independent evaluation pending — Fresh release (2026-08-28); no independent no-tools measurement published yet.
- MMLU-Pro: independent evaluation pending — Fresh release; vals.ai / Artificial Analysis runs pending.
- MMMU-Pro: not reported by the vendor — Text-only preview.
Architecture and hardware
- Parameters
- 770B total, 49B active per token (MoE)
- Estimated VRAM at Q4
- ~443 GB, Frontier class
- Quantization formats
- FP8
- Recommended runtime
- vLLM
- License
- Apache 2.0
Links
More from Tencent
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.