CL-SN55 941 —
GPT-61S 890 —
GPT-56T 861 —
GPT-6S 837 -1.2%
CL-OP55X 831 -1.4%
MUSE-SPK 804 —
QWEN-38X 795 —
GPT-6A 782 —
CL-OP5H 724 -1.6%
CL-OP5X 722 -1.6%
KIMI-K3X 718 -0.1%
CL-OP46H 707 -1.7%
CL-FAB5H 705 -7.4%
CL-OP55H 705 -1.7%
CL-OP47H 696 -1.7%
GEM-37FH 661 -9.7%
GEM-38FH 656 -4.8%
GPT-56S 580 -0.5%
CL-OP47 545 -0.5%
INKL 531 —
GEM-3P 494 -0.2%
CL-OP46 491 —
GEM-31P 486 -0.4%
CL-OP48 484 —
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 890 —
GPT-56T 861 —
GPT-6S 837 -1.2%
CL-OP55X 831 -1.4%
MUSE-SPK 804 —
QWEN-38X 795 —
GPT-6A 782 —
CL-OP5H 724 -1.6%
CL-OP5X 722 -1.6%
KIMI-K3X 718 -0.1%
CL-OP46H 707 -1.7%
CL-FAB5H 705 -7.4%
CL-OP55H 705 -1.7%
CL-OP47H 696 -1.7%
GEM-37FH 661 -9.7%
GEM-38FH 656 -4.8%
GPT-56S 580 -0.5%
CL-OP47 545 -0.5%
INKL 531 —
GEM-3P 494 -0.2%
CL-OP46 491 —
GEM-31P 486 -0.4%
CL-OP48 484 —
GPT-6L 436 —
← Back to feed

Tencent Open-Sources Hy3-Preview: 295B MoE at 74.4% SWE-Bench Verified With Only 21B Active Params

Tencent open-sourced Hy3-preview on April 23, the first model out of a complete rebuild of the Hunyuan training infrastructure. At 295B total parameters with 21B active per forward pass, it is a more parameter-efficient architecture than its open-weight peers — DeepSeek-V3 runs 37B active, Kimi K2 runs 32B active — while matching or beating both on several key benchmarks.

The release is also Yao Shunyu’s first public deliverable since joining Tencent as Chief AI Scientist and Head of LLM. From training start to launch took under three months.

Benchmark Numbers

On agentic coding and search benchmarks, which Tencent says saw the largest gains from this rebuild:

BenchmarkHy3-preview
SWE-bench Verified74.4%
Terminal-Bench 2.054.4%
BrowseComp67.1%
WideSearch70.2%

That puts Hy3-preview ahead of GLM-5.1 and Kimi K2.5 on search-agent tasks, and closing meaningful ground on Claude Opus 4.6 (80.8%) and GPT-5.4 (78.6%) on SWE-bench Verified, from a model with less than a third of their provider’s compute footprint.

On base model evals, Hy3-preview outperforms DeepSeek-V3 on math (MATH: 76.28 vs 59.37, GSM8K: 95.37 vs 88.15) and multilingual tasks (MMMLU: 80.15 vs 79.54) despite activating 43% fewer parameters per token. It trails Kimi K2 on some English tasks (SimpleQA: 26.47 vs 34.37) and Chinese benchmarks.

Architecture

Hy3-preview is a decoder-only transformer with 80 layers, 64 attention heads (GQA with 8 KV heads), 192 experts with top-8 activation, and a 256K context window. A single Multi-Token Prediction layer adds 3.8B parameters for inference latency improvements. The model operates in BF16 and requires H20-3e class GPUs or equivalent to serve on 8 GPUs.

The Infrastructure Angle

Tencent frames this release as the start of an era rather than a product milestone. The rebuilt training infrastructure — which they describe as a “full-link reconstruction” — is what they expect to compound. Hy3-preview is explicitly described as the beginning, not the destination.

Internal evaluation shows a 55-56% win rate over prior Hunyuan models in blind user evaluation. On production deployments (CodeBuddy, WorkBuddy), the team reports 54% reduction in first-token latency and 47% reduction in end-to-end duration versus Hy2, with 99.99% success rates.

Availability

Weights are on Hugging Face, ModelScope, and GitCode as Tencent-Hunyuan/Hy3-preview (instruct) and Tencent-Hunyuan/Hy3-preview-Base. The license is Tencent Hy Community License Agreement — not Apache 2.0 or MIT. Commercial use terms require review before deployment.

Hy3-preview is now appearing on Artificial Analysis comparisons alongside GPT-5.5 (high) as a reference point for open-weight reasoning capability.