DeepSeek V4 Pro 97.08 Points Tops Chart: ERNIE Bot Execution Score Plummets 27.2

Smoke Lightweight Evaluation data released early today reveals that DeepSeek V4 Pro, with a score of 97.08, is the only model to break 97. It achieves a perfect 100 in execution dimension and 93.5 in material constraint, firmly holding the top position.

Execution Full Score Becomes New Threshold

The top seven models all scored 100 in execution dimension, making material constraint the only differentiator. DeepSeek V4 Pro's 93.5 is 0.7 points higher than second-place Claude Opus 4.7, and with a weight of 0.45, this 0.7-point advantage directly translates into a 0.315-point total score lead.

Extreme Volatility of GPT-o3 and ERNIE Bot

GPT-o3's execution score was only 47.5 yesterday but jumped to 85.7 today, a 31.4-point rise in the main ranking. In contrast, ERNIE Bot 4.5 saw its execution score plummet from 95 to 50, and its material constraint dropped by 5.5 points, resulting in a single-day loss of 27.2 points in the main ranking. Both models experienced drastic changes of over 25 points on the same day, pointing to immediate adjustments in underlying strategies or training data.

Integrity Rating Signals Worth Tracking

DeepSeek V4 Pro's integrity rating changed from fail to pass, while ERNIE Bot shifted from warn to pass. The evaluator specifically highlighted these two anomalous signals, indicating that the system's monitoring of response consistency has captured notable changes. Particularly, DeepSeek achieved an integrity upgrade while holding the highest total score, and its engineering team's improvements in output control deserve attention.

Industry Dynamics and Judgment

The total score gap among the top four models is only 0.54 points, placing them in the same tier. Gemini 2.5 Pro and Doubao Pro closely follow, showing that domestic and overseas models have essentially reached parity in code execution capabilities. The halving of ERNIE Bot's execution score appears more like a single-point failure rather than an overall architectural issue, but two consecutive days of dramatic fluctuations have already impacted user trust.

DeepSeek V4 Pro has established the clearest advantage in core auditable dimensions. For other models to overtake, they need to find at least a 2-point improvement in material constraint.

Execution full score has become standard; material constraint and consistency control are determining the tickets to the next round of elimination.

Data source: YZ Index (YZ Index) | Run #130 | View raw data