Successor to ERNIE 5.0, launched at Baidu Create 2026. Derived from ERNIE 5.0 via elastic sub-network extraction — total params ~1/3 of 5.0, active params ~1/2 — with a disaggregated fully-asynchronous RL infrastructure and a four-stage post-training pipeline including Multi-Teacher On-Policy Distillation. Baidu claims pre-training cost is ~6% of comparable industry peers.

Self-reported benchmarks: AIME26 99.6 (with tool use), LMArena Arena Search 1223 (#4 globally, #1 among Chinese models), surpasses DeepSeek-V4-Pro on τ³-bench and SpreadsheetBench-Verified, approaches closed-source frontier on GPQA Diamond and MMLU-Pro. Not yet scored on Artificial Analysis.

Model Details

Architecture MOE

Benchmark Scores

Benchmark Score Mode
AIME 2026 99.6 with tool use
LMArena Arena Search 1223 #4 global, #1 Chinese
frontierreasoningmoeagentic

Related