Modèle Updated 2026-04
SLM (Small Language Model)
Small Language Model
Definition
An SLM is a compact language model optimized to run on local devices with targeted performance on specific tasks.
See also in the glossary
L
LLM (Large Language Model)
An LLM is an AI model trained on billions of texts, capable of understanding and generating human language.
F
Foundation Model
A foundation model is a large AI model pre-trained on massive data, adaptable to multiple tasks.
Q
Quantization
Quantization reduces the precision of numbers in an AI model to make it smaller and faster, with minimal quality loss.
A
AI Inference
Inference is the process of using a trained AI model to generate predictions or responses from new data.
Tools that use slm
Frequently Asked Questions
What's the difference between SLM and LLM?
An LLM has billions of parameters and requires powerful GPUs. An SLM has a few billion parameters and can run on a laptop or smartphone.
Are SLMs worse?
On general tasks, yes. But on specific tasks they're optimized for, they can rival much larger LLMs.
What is a small language model (SLM)?
An SLM is a compact language model, typically from a few hundred million to around 10 billion parameters, designed to run efficiently on laptops, phones, or edge devices. Examples include Microsoft's Phi-3, Google's Gemma, Llama 3.2 1B/3B, and Mistral small models. Unlike large LLMs that need data-center GPUs, SLMs trade broad general knowledge for lower cost, faster latency, and on-device privacy.
Are SLMs cheaper to run than LLMs?
Yes, considerably. With far fewer parameters, SLMs need less memory and compute, so they cost less per token and often run on a single consumer GPU, a CPU, or even a phone, with no API fees. A model like Phi-3-mini or Gemma 2B can serve many tasks locally where a frontier LLM would require expensive cloud inference. The trade-off is reduced general capability.
Which is the best small language model?
There's no single winner; it depends on the task and hardware. Microsoft's Phi-3 punches above its size on reasoning, Google's Gemma 2 is strong and openly licensed, and Llama 3.2 1B/3B and Qwen2.5 small variants are popular for on-device use. For tight memory budgets, TinyLlama or sub-2B models work. Benchmark a few candidates on your actual workload rather than trusting leaderboards alone.
Is Llama an SLM or an LLM?
Both, depending on the variant. "Llama" is Meta's model family (the name stands for Large Language Model Meta AI), and its big versions like Llama 3 70B or 405B are full LLMs. But smaller releases such as Llama 3.2 1B and 3B are effectively SLMs, built to run on phones and edge devices. TinyLlama, a separate community project, is firmly in SLM territory at 1.1B parameters.
Is DeepSeek an LLM or an SLM?
DeepSeek is primarily an LLM developer. Its flagship models, such as DeepSeek-V3 and the R1 reasoning model, are very large, with hundreds of billions of parameters. However, DeepSeek also ships smaller distilled versions, for example R1 distilled into 1.5B to 14B variants based on Qwen and Llama, which fall into SLM territory and can run on modest local hardware.
How do you turn an LLM into an SLM?
Several compression techniques shrink a large model into a small one. Knowledge distillation trains a compact "student" to mimic a large "teacher", as DeepSeek did with its R1-distilled models. Quantization lowers numerical precision (for example to 4-bit) to cut memory, while pruning removes redundant weights. The goal is to keep most task performance while making the model small enough for on-device, low-latency inference.
What is the future of SLMs?
SLMs are central to the shift toward on-device and agentic AI. As phones and laptops gain dedicated NPUs, more inference moves local for privacy, offline use, and lower cost. Expect smaller models fine-tuned for narrow tasks, used as fast, cheap workers inside agent pipelines, with large frontier LLMs reserved for the hardest reasoning. The 2026 trend is many specialized small models rather than one giant model for everything.
