Small Language Models (SLM)

Models under about 10B parameters that fit consumer devices.

Tier Grades

Tier A: Frontier AI labs

feasible

high confidence

Microsoft, Hugging Face, Alibaba, OpenAI and NVIDIA are observed producing SLMs.

Tier B: Indonesian AI labs

feasible

medium confidence

Training from scratch needs about 384 H100 for 24 days (SmolLM3), below the 70B scale b reached with Sahabat-AI, though the notes verify no Indonesian national compute programme, BRIN facility or ITB GPU inventory (T11 Part E item 2).

Tier C: Institution units such as ITB

stretch

high confidence

From-scratch training at 384 H100 for 24 days exceeds the academic 1 to 8 GPUs, while fine-tuning at a few GPU-hours fits.

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

feasible

high confidence

Fine-tuning needs a few GPU-hours and serving is 10 to 30x cheaper than large models, which suits proof-of-concept delivery (not from-scratch training).

Where this comes from

Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.

Anchors

ComputeSmolLM3 3B: 384 H100 for 24 days from scratch; fine-tuning an SLM needs a few GPU-hours; serving a 7B model is 10 to 30x cheaper than 70B to 175B.
DataPhi-3-mini 3.3T tokens; SmolLM3 about 11T tokens.
Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): NVIDIA position paper has 8 authors.
Time horizonMonths.
Shelf-life signals40 to 70% of LLM calls in three open agent systems could be replaced by SLMs (NVIDIA case-study estimate)
Pipeline roleTier A and open labs release SLMs; Tier B and Tier C fine-tune them for local languages and tasks; Tier D deploys them on devices for proof of value.

Where this comes from

Team size is a proxy (author counts of exemplar papers), not a headcount.

Who and What

Exemplars

Phi-3

SmolLM3Qwen3 small OpenAI

gpt-oss-20b

Observed actors

Microsoft AI Hugging Face

Hugging Face Alibaba (Qwen)

Alibaba (Qwen) OpenAI

OpenAI

NVIDIA

Linked models

+132

Tiny Aya globaledge LLMproduced by

Tiny Aya wateredge LLMproduced by

Qwen3.5-0.8Bmultimodal LLMproduced by

Qwen3.5-2Bmultimodal LLMproduced by

Qwen3.5-4Bmultimodal LLMproduced by

Qwen3.5-9Bmultimodal LLMproduced by

Qwen3.5 small (0.8B, 2B, 4B, 9B)multimodal LLMproduced by

Nemotron 3 Nano 4BLLMproduced by

Falcon-H1R-7Breasoningproduced by

Phi-4-miniedge LLMproduced by

Ministral 3 14B Instruct (25.12)multimodal LLMproduced by

Jamba Reasoning 3Breasoningproduced by

Granite-4.0-H-Microedge LLMproduced by

VaultGemma 1BLLMproduced by

OpenBMB (Tsinghua NLP Lab and ModelBest)

MiniCPM4.1 8Bedge LLMproduced by

FastVLM 0.5Bmultimodal LLMproduced by

FastVLM 7Bmultimodal LLMproduced by

NVIDIA Nemotron Nano 9B v2reasoningproduced by

GPT-5 nanoreasoningproduced by

Gemma 3 270Medge LLMproduced by

Teuken-7B v0.6LLMproduced by

Apple on-device foundation model (2025)edge LLMproduced by

SmolLM3-3Bedge LLMproduced by

Phi-4-mini-flash-reasoningreasoningproduced by

Gemma 3n E2Bedge LLMproduced by

MiniCPM4 / 4.1edge LLMproduced by

MiniCPM4 0.5Bedge LLMproduced by

MiniCPM4 8Bedge LLMproduced by

Gemma 3n E4Bedge LLMproduced by

Qwen3-0.6B-BaseLLMproduced by

Qwen3-4B-BaseLLMproduced by

Qwen3-0.6BLLMproduced by

Qwen3-1.7BLLMproduced by

Qwen3-4BLLMproduced by

Qwen3-8BLLMproduced by

Granite 3.3 8B InstructLLMproduced by

Llama-3.1-Nemotron-Nano-8B-v1reasoningproduced by

Gemma 3edge LLMproduced by

Gemma 3 4Bmultimodal LLMproduced by

Gemma-3-4B-itLLMproduced by

Phi-4-mini-instructedge LLMproduced by

Granite 3.2 8B InstructLLMproduced by

ALLaM-7B-Instruct-previewLLMproduced by

SmolVLM2 256M Video Instructmultimodal LLMproduced by

SmolVLM2 500M Video Instructmultimodal LLMproduced by

SmolVLM2 2.2B Instructmultimodal LLMproduced by

OpenThinker-7Breasoningproduced by

DeepSeek-R1-Distill-Llama-8Breasoningproduced by

DeepSeek-R1-Distill-Qwen-1.5Breasoningproduced by

DeepSeek-R1-Distill-Qwen-7Breasoningproduced by

InternLM3 8B instructLLMproduced by

Olmo 3 7BLLMproduced by

Phi-4-multimodal-instructmultimodal LLMproduced by

FastVLMmultimodal LLMproduced by

Granite 3.1 8B InstructLLMproduced by

Falcon3-7B-InstructLLMproduced by

SmolVLMmultimodal LLMproduced by

SmolVLM Instructmultimodal LLMproduced by

MobileLLM 1Bedge LLMproduced by

MobileLLM 125Medge LLMproduced by

Llama 3.2 1Bedge LLMproduced by

Llama 3.2 3Bedge LLMproduced by

Granite 3LLMproduced by

Granite 3.0 8B InstructLLMproduced by

Gemma 2 2B JPN ITLLMproduced by

Qwen2.5-0.5B-InstructLLMproduced by

Qwen2.5-7B-InstructLLMproduced by

Qwen2.5 0.5BLLMproduced by

Nemotron-Mini-4B-Instructedge LLMproduced by LLama-3-8B-Tele-itLLMproduced by

Phi-3.5-vision-instructmultimodal LLMproduced by

Llama-3.1-8B-InstructLLMproduced by

Gemma 2 2BLLMproduced by

InternLM2.5 7B chatLLMproduced by

Llama-3-ELYZA-JP-8BLLMproduced by

Gemma 2 9BLLMproduced by

GLM-4-9B-ChatLLMproduced by

Yi-1.5 9B ChatLLMproduced by

OpenELMedge LLMproduced by

OpenELM 1.1B Instructedge LLMproduced by

OpenELM 270M Instructedge LLMproduced by

OpenELM 3B Instructedge LLMproduced by

Komodo-7BLLMproduced by

MobileLLMedge LLMproduced by

BioMistral 7BLLMproduced by

InternLM2 chat 7BLLMproduced by

Phi-2edge LLMproduced by

Swallow 7B InstructLLMproduced by

ChatGLM3-6BLLMproduced by

ChatGLMLLMproduced by

ChatGLM2-6BLLMproduced by

Falcon-RW-1BLLMproduced by

Falcon-7B-InstructLLMproduced by

Falcon-7BLLMproduced by

ERNIE-4.5-0.3BLLMproduced by

FunctionGemma 270Medge LLMproduced by

Gemini Nanoedge LLMproduced by

Gemma2-2b-ITLLMproduced by

Gemma2-9b-ITLLMproduced by

Gemma 3 1Bedge LLMproduced by

Gemma 3nedge LLMproduced by

SmolLM2edge LLMproduced by

SmolLM3-3B-BaseLLMproduced by

granite-4.1-8bLLMproduced by

Granite-4.1-8B-BaseLLMproduced by

Granite 4 350Medge LLMproduced by

LFM2.5-Thinking 1.2Bedge LLMproduced by

Llama 3.1 8BLLMproduced by

Llama-3.2-1B-InstructLLMproduced by

Llama 3.2 1B quantized (QLoRA)edge LLMproduced by

Llama 3.2 1B quantized (SpinQuant)edge LLMproduced by

Llama 3.2 3B quantized (QLoRA)edge LLMproduced by

Llama 3.2 3B quantized (SpinQuant)edge LLMproduced by

Llama 3 8BLLMproduced by

Meta-Llama-3-8B-InstructLLMproduced by

Phiedge LLMproduced by

Phi-3.5-miniLLMproduced by

phi-4LLMproduced by

Phi-4-mini-reasoningreasoningproduced by

Ministral-3-3B-Base-2512LLMproduced by

Mistral 7BLLMproduced by Mistral-7B-v0.1LLMproduced by

Mistral-7B-v0.3LLMproduced by Moondream 1.8Bedge LLMproduced by

Ministral 3B (24.10)edge LLMproduced by

BitCPM4 0.5Bedge LLMproduced by

BitCPM4 1Bedge LLMproduced by

MiniCPM-V 4.6 1.3Bmultimodal LLMproduced by

MiniCPM3-1Bedge LLMproduced by

Falcon-E 3BLLMproduced by

Falcon-H1-Tinyedge LLMproduced by

Singapore University of Technology and Design

TinyLlama 1.1Bedge LLMproduced by

Papers

SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language ModelLoubna Ben Allal et al., arXiv preprint, 2025 state of the art
Phi-3 Technical Report: A Highly Capable Language Model Locally on Your PhoneMarah Abdin et al., arXiv preprint, 2024 state of the art
Gemma 2: Improving Open Language Models at a Practical SizeGemma Team et al., arXiv preprint, 2024 state of the art
MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use CasesZechun Liu et al., ICML 2024 (per arXiv comment), 2024 state of the art
A Survey of Small Language ModelsChien Van Nguyen et al., arXiv preprint, 2024 survey

Where this comes from

The papers to read first: where the direction began, its current best, and a survey when one exists.

Sources

Qwen3 Technical Report arxiv.org https://arxiv.org/abs/2505.09388
Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone arxiv.org https://arxiv.org/abs/2404.14219
SmolLM3: smol, multilingual, long-context reasoner huggingface.co https://huggingface.co/blog/smollm3
Small Language Models are the Future of Agentic AI arxiv.org https://arxiv.org/html/2506.02153v1
openai.com/index/introducing-gpt-oss https://openai.com/index/introducing-gpt-oss/

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Small Language Models (SLM)", https://aurion.semai.id/topics/small-language-models/, records of October 2026.

Small Language Models (SLM)

Tier Grades

Tier A: Frontier AI labs

Tier B: Indonesian AI labs

Tier C: Institution units such as ITB

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

Anchors

Who and What

Links

Models 156

Papers

Sources