DEV.co
Open-Source LLM · HuggingFaceTB

SmolLM2-360M

SmolLM2-360M is a 361M-parameter open-source language model from HuggingFace designed for on-device inference. It trades raw capability for efficiency, suitable for edge deployment and resource-constrained environments. Apache 2.0 licensed and non-gated, making it freely usable and modifiable.

Source: HuggingFace — huggingface.co/HuggingFaceTB/SmolLM2-360M
362M
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
90.5k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperHuggingFaceTB
Parameters362M
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads90.5k
Likes112
Last updated2025-02-06
SourceHuggingFaceTB/SmolLM2-360M

What SmolLM2-360M is

Transformer decoder architecture pretrained on 4 trillion tokens (FineWeb-Edu, DCLM, The Stack, plus curated datasets) in bfloat16 precision on 128 H100 GPUs using nanotron. Instruct variant fine-tuned via supervised fine-tuning and Direct Preference Optimization. Memory footprint ~724 MB at full precision; supports bfloat16 and quantization. Evaluated via lighteval on zero-shot benchmarks.

Quickstart

Run SmolLM2-360M locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="HuggingFaceTB/SmolLM2-360M")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Edge and mobile deployment

360M parameters and ~724 MB memory footprint enable deployment on constrained devices (mobile, embedded systems, edge servers) without cloud dependency.

Lightweight instruction-following assistants

Instruct variant shows improved performance on IFEval (41.0%) compared to predecessor. Suitable for domain-specific chatbots and automation when inference latency/cost matters more than raw accuracy.

Local privacy-critical applications

Non-gated, Apache 2.0 licensed model can run entirely on-premise without external API calls, enabling sensitive document processing, customer support, or compliance-heavy workflows.

Running & fine-tuning it

Minimum ~1–2 GB GPU VRAM for inference at bfloat16; full precision loads ~724 MB model weights. CPU inference possible but slower. For fine-tuning: LoRA/QLoRA feasible on 8–16 GB VRAM; full fine-tuning likely requires 24+ GB or distributed setup. Estimate only; validate on target hardware.

Model supports standard transformer fine-tuning. Card reports SFT and DPO applied for instruct variant, indicating standard PyTorch + HF Trainer compatibility. LoRA/QLoRA strongly recommended to reduce compute (360M base is small; full fine-tuning overhead may outweigh benefit on small datasets). Accelerate library compatible. No instruction-time preference for specific fine-tuning framework documented.

When to avoid it — and what to weigh

  • High-accuracy knowledge tasks — GSM8K (5-shot) scores only 7.43% vs Qwen2.5-0.5B's 26.8%. TriviaQA at 16.9%. Not suitable for fact-heavy applications requiring strong reasoning or retrieval.
  • Long-context or multi-turn dialogue — Context length is Unknown in provided data. No evidence of long-context capability. If your application requires deep conversation history or document processing, investigate context window before deployment.
  • Non-English or multilingual tasks — Model card explicitly states primary understanding of English. Multilingual or non-English generation quality is not documented and should be independently validated.
  • Real-time factuality requirements — Card notes generated content may not be factually accurate or logically consistent. Unsuitable for tasks where hallucination is unacceptable (medical diagnosis, legal advice, financial recommendations).

License & commercial use

Apache License 2.0: permissive OSI-approved license allowing commercial use, modification, and redistribution under same terms. No additional restrictions noted.

Apache 2.0 is a permissive OSI license explicitly permitting commercial use, including proprietary applications. No gating or restrictions on commercial deployment. You may use, modify, and distribute SmolLM2-360M commercially without permission or royalty. Attribution recommended but not legally required. Confirm compliance with derived dataset licenses (FineWeb-Edu, DCLM, The Stack, UltraFeedback) if you republish.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityLow
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

No explicit security audit or adversarial robustness testing documented. Standard LLM risks apply: model may perpetuate biases from training data (FineWeb-Edu, DCLM, Stack, curated datasets); no guarantees of robustness to prompt injection or jailbreaking. On-device deployment eliminates data transmission risk. Always validate outputs in sensitive use cases. No mention of differential privacy or data leakage mitigations.

Alternatives to consider

Qwen2.5-0.5B

Similar 500M scale, stronger on reasoning (GSM8K 33.4% vs 7.43%) and MT-Bench (4.16 vs 3.66), but marginally weaker on instruction-following (IFEval 31.6% vs 41.0%). Comparable licensing (check Qwen's license separately).

SmolLM-360M (predecessor)

Same parameter count, older version with lower instruction-following (IFEval 19.8%) and reasoning scores. Evaluate only if targeting legacy compatibility.

Phi-3-mini (3.8B)

Larger (~10x parameters) but still efficient; likely stronger reasoning and knowledge. Trade-off: higher memory, slower on edge. Consider if accuracy trumps deployment constraints.

Software development agency

Ship SmolLM2-360M with senior software developers

Start with a proof-of-concept on your hardware. Clone from HuggingFace, run the provided examples, and benchmark against your use case. For production edge deployment or fine-tuning, consult our custom LLM apps and private deployment services.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

SmolLM2-360M FAQ

Can I use SmolLM2-360M commercially?
Yes. Apache 2.0 is a permissive OSI license permitting commercial use without restriction or royalty. Ensure you comply with licenses of the upstream datasets (FineWeb-Edu, DCLM, The Stack, UltraFeedback) if you republish or redistribute.
What GPU or hardware do I need?
For inference: 1–2 GB VRAM (GPU optional; CPU viable but slow). For fine-tuning: 8–16 GB VRAM with LoRA/QLoRA recommended, or 24+ GB for full fine-tuning. Model card shows ~724 MB memory footprint at full precision. Test on your target hardware before production.
How does it compare to Qwen2.5-0.5B?
SmolLM2-360M excels at instruction-following (IFEval 41.0% vs 31.6%) and commonsense reasoning. Qwen2.5-0.5B is stronger at math/reasoning (GSM8K 33.4% vs 7.43%). No clear winner; choose based on your task.
Is the model card for the instruct or base version?
Card documents both. Base model (pretrained only) and instruct variant (fine-tuned with SFT + DPO) are available. Evaluation tables show results for both; instruct version typically better for chat and instruction-following tasks.

Custom software development services

DEV.co helps companies turn open-source tools like SmolLM2-360M into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to Deploy SmolLM2-360M?

Start with a proof-of-concept on your hardware. Clone from HuggingFace, run the provided examples, and benchmark against your use case. For production edge deployment or fine-tuning, consult our custom LLM apps and private deployment services.