SmolLM3-3B: benchmark results

Hugging Face's fully open Apache-2.0 3B model (July 2025) with dual think/no-think reasoning modes, 128K context, and a published training recipe and data mix. Provider: HuggingFace. Released 2025-07-08. Access: Open.

Aggregate Rating 1392 ± 13, rank #1252 of 1656 rated models, from 408 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
RoBench-25 (Awareness Instruction) - No Trigger1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness Instruction) - Trigger A1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness Instruction) - Trigger B1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness Instruction) - Trigger C1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness Instruction) - Trigger D1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness Instruction) - Trigger E1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness Instruction) - Trigger F1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness Instruction) - Trigger G1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness and Recovery Instructions) - No Trigger1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness and Recovery Instructions) - Trigger A1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness and Recovery Instructions) - Trigger B1Clipped Heidke skill score (0-1; chance-corrected accuracy o100
RoBench-25 (Awareness and Recovery Instructions) - Trigger D1Clipped Heidke skill score (0-1; chance-corrected accuracy o100

Interactive version: theaggregate.ai/model?slug=smollm3-3b · How It Works · Data refreshed daily, snapshot 2026-10-20.