One-line: Tencent’s frontier-scale 770B/49B-active MoE flagship (Gated DSA + MLA, 1M context) with strong agentic-coding numbers — ⚠ large, not local-runnable, but at the open-source coding frontier.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
—
—
MBPP
—
—
LiveCodeBench
—
—
Aider polyglot
—
—
SWE-bench Verified
—
—
SWE-bench Multilingual
82.9
lab (Tencent benchmark appendix, lab-reported)
SWE-bench Pro
65.7
lab (Tencent benchmark appendix, lab-reported)
Terminal-Bench 2.1
70.8
lab (Tencent benchmark appendix, lab-reported)
DeepSWE
64.3
lab (Tencent benchmark appendix, lab-reported)
NL2Repo-Bench
43.8
lab (Tencent benchmark appendix, lab-reported)
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
—
τ-Bench
—
—
ToolACE
—
—
MCP-Atlas
73.0
lab (Tencent benchmark appendix, lab-reported)
Toolathlon-Verified
56.2
lab (Tencent benchmark appendix, lab-reported)
WideSearch
81.9
lab (Tencent benchmark appendix, lab-reported)
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
90.9
lab (Tencent benchmark appendix, lab-reported)
MATH
—
—
HLE (no tools, text-only)
34.4
lab
HLE (with tools, text-only)
51.9
lab
Efficiency
Active params: 49B active per token (770B total; 256 routed experts, top-8 + 1 shared)
Context length: 1,048,576 tokens (1M)
VRAM (fp16 rough): ~98 GB for active weights alone; full model needs 8×80GB GPUs (README ships tp-size-8 recipes) — ⚠ not local-runnable
Architecture: MoE with Gated DeepSeek Sparse Attention (DSA) + MLA, IndexCache cross-layer sparse index reuse, iHC residual stream; 78 layers; 1 native MTP layer (10B/0.7B active) for speculative decoding
Also available: Hy4-preview-FP8 (quantized), finetuning pipeline, AngelSlim compression toolkit
What makes it notable
Hy4 preview is Tencent’s flagship open-weight MoE and claims to sit at the open-source frontier for long-horizon software engineering. Its coding/agentic numbers are strong for the open-weight space — Terminal-Bench 2.1 70.8, SWE-bench Multilingual 82.9, MCP-Atlas 73.0 — and it positions itself slightly ahead of GLM 5.3 and Kimi K3 in a blind internal expert eval (163 experts, 203 engineering tasks). ⚠ Large (>70B): 770B total / 49B active, requiring 8×80GB GPUs — not for local/agent-loop deployment on consumer/homelab hardware. ⚠ Benchmarks are lab-reported from Tencent’s own image appendix (no SWE-bench Verified, Aider, or BFCL published); an independent leaderboard (BenchLM) verified SWE-bench Pro 65.7 and SWE-Multilingual 82.9 against that same appendix. Early/preview build: it over-reasons on complex tasks and over-verifies its own work.