文心大模型5.1预览版
The Wenxin Model 5.1 Preview is a large-scale reasoning model developed by Baidu. It boasts an extremely massive scale of approximately 8 trillion parameters and is a foundation model with a 128K context window.
Parameters
8000.0B
Context Window
128K
License
Proprietary
Release Date
2026-04-30
API Pricing
API pricing for this model is not yet available
Strengths
- ・Overwhelming ~8 trillion parameters
- ・Advanced reasoning capabilities
- ・Long 128K context window
Weaknesses
- ・Closed licensing system
- ・Non-public model architecture
- ・Instability as a preview version
Use Cases
- ・Complex logical reasoning tasks
- ・Analysis of large documents
- ・Building advanced knowledge bases
Deep Analysis
Architecture
MoE (~800B total)
Extracted from ERNIE 5.0 (2.4T) sub-model matrix
Active Parameters
~400B (half of total)
Compressed from ERNIE 5.0
LMArena Search Arena
#4 globally (1223 score)
#1 among Chinese models
AIME 2026 (with tools)
99.6
Second only to Gemini 3.1 Pro
Training Cost
~6% of comparable models
Extreme efficiency
Release Date
May 8, 2026
Latest flagship
LMArena Text
#14 globally
Highest Chinese model position
Strengths
- ・#4 globally on LMArena Search Arena
- ・AIME26 score of 99.6 (near-perfect)
- ・Only 6% pre-training cost of comparable models
- ・Strong agentic capabilities (surpasses DeepSeek-V4-Pro on tau3-bench)
- ・Top Chinese model on multiple leaderboards
- ・Approaches Gemini 3.1 Pro in creative writing
Weaknesses
- ・Very large model, limited self-hosting options
- ・Chinese model with English as secondary
- ・Relatively new (May 2026), limited production track record
Competitor Comparison
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| Gemini 3.1 Pro | - | - | - | Paid |
| Claude Opus 4.6 | - | - | - | Paid |
| DeepSeek V4 Pro | - | - | - | Lower |
ERNIE 5.1 is Baidu's latest flagship model released May 8, 2026. With ~800B total parameters, it achieves #4 globally on LMArena Search Arena, AIME26 score of 99.6, and surpasses DeepSeek-V4-Pro on agentic benchmarks, all while using only 6% of the pre-training cost of comparable models.
Sources
Analysis generated: 2026-05-24