Model
Model explorer

Hunyuan-TurboS

CLOSED
Tencent · Hunyuan family · released Feb 27, 2025

First industry-deployed ultra-large Hybrid Transformer-Mamba MoE (560B/56B-active) fast-thinking model with adaptive CoT; ranked top-7 on LMSYS Chatbot Arena (1356).

ReasoningCodingVisionFunction callingTool useAgentic
1306.0
Elo · rank #154
Parameters
560B
Active params
56B (MoE)
Context
256K tokens
Architecture
Hybrid Transformer-Mamba2 MoE, 128 layers with AMF/MF block pattern (560B total / 56B active)
License
Proprietary
Languages
API price (in/out)
$0.12 / $0.3
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AIMEMath40.0%#125
best: GPT-5.2 · 100.0%
AlpacaEvalHuman preference76.0%#5
best: Tülu 2+DPO 70B · 95.1%
Arena-HardHuman preference91.9%#7
best: Qwen3-235B-A22B (Non-Thinking) · 96.1%
Arena EloHuman preference1356#29
best: Claude Fable 5 · 1505
BIG-Bench HardReasoning90.8%#5
best: ERNIE 4.5 300B-A47B · 94.3%
C-EvalKnowledge88.1%#17
best: Qwen3.6-Plus · 93.3%
CMMLUKnowledge89.4%#5
best: Doubao-1.5-Pro · 90.9%
DROPReasoning89.8#8
best: Hunyuan-T1 · 93.1
GPQA DiamondReasoning62.1%#156
best: GPT-5.6 · 94.6%
GSM8KMath94.4%#17
best: Llama 3.1 405B · 96.8%
HumanEvalCoding89.0%#24
best: Claude Opus 4.5 · 99.4%
IFEvalReasoning87.6%#38
best: Gemma 4 26B A4B · 98.5%
LiveCodeBenchCoding43.0%#130
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MATH-500Math90.0%#47
best: GPT-5 · 99.4%
MMLU-ProKnowledge78.0%#66
best: Claude Fable 5 · 91.5%
MMLUKnowledge85.8%#52
best: OpenAI o3 · 92.9%
MT-BenchHuman preference9.3#4
best: Hunyuan-Large (A52B) · 9.4
Run it locally
Closed weights — available via API only. No local deployment.
Input / M tok
$0.12
Output / M tok
$0.3
API price $0.12/$0.3 · each benchmark row carries its own source badge (see methodology)