GGUF IQ4_XS (qtum)
source ↗Perplexity
3.08+0.14baseline 2.94
Hy4-preview
- Organization
- Tencent
- Parameters
- 770B total / 49B active (+10B MTP)
- Context
- 1M
- License
- Apache-2.0
- Release
- 2026-08-27
HYV4ForCausalLM, 78 layers (77 MoE, 256 routed + 1 shared, top-8), Gated DSA attention + IndexCache, iHC 4 streams
Tencent's Hy4-preview is a 770B-total / 49B-active MoE with a 10B MTP module, Gated DSA attention, a 1M context and an Apache-2.0 license. The qtum GGUF builds and the Oxmiq NVFP4 W4A16 card publish per-quant perplexity and benchmark tables, while the AngelSlim STQ1_0 GGUF (213.66 GiB) is the only checkpoint with measured serving numbers, at 204.56 t/s prefill and 20.47 t/s decode on 8x H20.
Hy4-preview is a 770B total / 49B active (+10B MTP) Apache-2.0 model from Tencent with 1M context. Quantization loss is not measured here — no published per-quant delta shows a quality drop versus the full-precision baseline.
Base-model benchmarks
Terminal-Bench 2.1
85.4GPQA Diamond
92.3SWE-bench Pro
65.7DeepSWE
64.3SWE-bench Multilingual
82.9Toolathlon Verified
74.1Apex Agents
37.1
Quant checkpoints
| Checkpoint | Family | Size | Source |
|---|---|---|---|
| tencent Hy4-preview BF16 ~727.4 GiB | BF16 | 727.4 GB | source ↗ |
| tencent Hy4-preview-FP8 ~813.8 GB | FP8 | 813.8 GB | source ↗ |
| AngelSlim Hy4-preview GGUF Q4_K_M GiB | GGUF | 435.2 GB | source ↗ |
| AngelSlim Hy4-preview GGUF UD-IQ1_M GiB | GGUF | 219.83 GB | source ↗ |
| AngelSlim Hy4-preview GGUF STQ1_0 GiB | GGUF | 213.66 GB | source ↗ |
| qtum Hy4-preview-GGUF IQ4_XS GiB | GGUF | 388.3 GB | source ↗ |
| qtum Hy4-preview-GGUF Q2_K GiB | GGUF | 269.6 GB | source ↗ |
| Oxmiq Hy4-preview-NVFP4-W4A16 ~490 GB NVFP4 W4A16 | NVFP4 | 490 GB | source ↗ |
Per-quant quality deltas
Serving metrics
What’s not published
- No official FP8-vs-BF16 accuracy comparison.
- The full benchmark appendix is image-only.
- AngelSlim speed numbers apply to STQ1_0 only.
Sources
FAQ
- What is Hy4-preview's license, context window and parameter count?
- Apache-2.0 license, 1M context window, 770B total / 49B active (+10B MTP) parameters (Tencent).
- How much quality does quantization cost Hy4-preview?
- Across 2 published deltas, quantized checkpoints differ from the full-precision baseline by 0.55 points on average (range 0.14 to 0.95).
- How many quant checkpoints are cataloged for Hy4-preview?
- 8 checkpoints across BF16, FP8, GGUF, NVFP4, sizes from 213.66 GB to 813.8 GB.
- When was Hy4-preview released?
- 2026-08-27.
- Where do the Hy4-preview benchmark numbers come from?
- From 5 official sources — vendor model cards, quantization publisher cards and serving recipe pages.