SLM
A language model with a parameter count typically in the range of several billion or fewer, designed for local deployment on edge devices. SLMs are optimized to run efficiently on consumer-grade hardware such as smartphones and laptops, representing a lightweight direction within the broader language model field.
Core Facts
Timeline (last 90 days)
The AI PC standard requires an NPU to locally run small language models (SLMs) with parameter counts in the range of 1 billion to 7 billion.
Small language models can be compressed to a few hundred MB or even smaller after quantization, making them small enough to store locally on a phone.
Small Language Models (SLMs) are lower cost, easier to manage for security purposes, and can run locally on edge devices.
A study published on arXiv constructed a benchmark around agent micro-tasks to examine whether off-the-shelf small language models (SLMs) are sufficient for these micro-tasks.
The architectural recommendation from the study is to place SLMs after a baseline that can reach the confidence-interval-supported threshold, and only invoke SLMs where the baseline cannot meet that standard.
In this learning framework, the SLM serves as the membership oracle and labeled data serves as the equivalence oracle.
The neurosymbolic router routes structured tasks (arithmetic, algebra, formal logic) to deterministic engines and keeps open-ended word problems for a small language model.
Forcing small language models to approximate deterministic problems probabilistically sacrifices accuracy and energy consumption, and much of the unreliability in edge small models is avoidable.
Many SLM (Spatial Language Model) datasets perform well on standard LiDAR metrics but see their scores drop significantly when entering real-world tilted scenarios.
Small language models (SLMs) typically have parameter counts within a few billion, aiming to run locally on devices such as phones and laptops.
All Facts (11)
Small Language Models (SLMs) generally refer to models with parameter counts between 1B and 7B.
90%UnverifiedThe AI PC standard requires an NPU to locally run small language models (SLMs) with parameter counts in the range of 1 billion to 7 billion.
50%UnverifiedSmall language models can be compressed to a few hundred MB or even smaller after quantization, making them small enough to store locally on a phone.
50%UnverifiedSmall Language Models (SLMs) are lower cost, easier to manage for security purposes, and can run locally on edge devices.
50%UnverifiedA study published on arXiv constructed a benchmark around agent micro-tasks to examine whether off-the-shelf small language models (SLMs) are sufficient for these micro-tasks.
50%UnverifiedThe architectural recommendation from the study is to place SLMs after a baseline that can reach the confidence-interval-supported threshold, and only invoke SLMs where the baseline cannot meet that standard.
50%UnverifiedForcing small language models to approximate deterministic problems probabilistically sacrifices accuracy and energy consumption, and much of the unreliability in edge small models is avoidable.
50%UnverifiedThe neurosymbolic router routes structured tasks (arithmetic, algebra, formal logic) to deterministic engines and keeps open-ended word problems for a small language model.
50%UnverifiedIn this learning framework, the SLM serves as the membership oracle and labeled data serves as the equivalence oracle.
50%UnverifiedMany SLM (Spatial Language Model) datasets perform well on standard LiDAR metrics but see their scores drop significantly when entering real-world tilted scenarios.
50%UnverifiedSmall language models (SLMs) typically have parameter counts within a few billion, aiming to run locally on devices such as phones and laptops.
50%