Atlas

Models

← All models

Nanbeige4-3B-Thinking-2511

Nanbeige LLM Lab · 2025-11-21 · 3.9B parameters

Nanbeige4-3B-Thinking-2511 is the reasoning-enhanced checkpoint in the Nanbeige4-3B family and an upgraded iteration of Nanbeige4-3B-Thinking-2510, improved through knowledge distillation and targeted reinforcement learning. Its base model was pretrained on a reported 23 trillion-token corpus, followed by post-training with supervised fine-tuning, distillation, and multi-stage reinforcement learning.

Benchmark scores

BenchmarkScore
BFCL v4 Overall Accuracy51.4
Capability122.2
Creative Writing v3 Elo949
Creative Writing v3 Rubric Score9.0
EQ-Bench 3675
EQ-Bench 3 (Rubric Score)38.1
EQ-Bench 3: Analytic7.5
EQ-Bench 3: Assertive6.8
EQ-Bench 3: Compliant4.1
EQ-Bench 3: Empathy3.7
EQ-Bench 3: Humanlike2.4
EQ-Bench 3: Insight5.3
EQ-Bench 3: Moralising7.9
EQ-Bench 3: Pragmatic2.3
EQ-Bench 3: Safety4.6
EQ-Bench 3: Social IQ2.0
EQ-Bench 3: Warm4.6
Longform Creative Writing31.5
Loading Atlas data…