Key takeaways

• The SmolLM2 360M Instruct is a compact (360M parameters) Transformer-based language model designed for creative text generation and basic instruction following.
• It excels in generating coherent and engaging creative content, especially when instruction-following demands are low.
• The model demonstrates significant weaknesses in complex constraint adherence, multi-step reasoning, and mathematical problem-solving, achieving 0% accuracy on the AIME 2025 benchmark.
• Common failure modes include inability to adhere to strict formatting/length guidelines, generating toxic/unreadable content, hallucinating facts, and producing lengthy but incorrect mathematical elaborations.
• While it can follow single, straightforward instructions with 28.68% accuracy on the IFEval benchmark, it struggles with more complex instructions and mathematical reasoning.

Technical specifications

Context length

8k tokens

Max output

—

Architecture

Transformer

Modalities

text->text

Parameters

~0.36B

Open weights

Yes

Licence

Apache 2.0

Regions

Independent

Vendor

Hugging Face

Provider model key

huggingface/smollm2-360m-instruct