Comparação de modelos muito pequenos de IA de pesos abertos (≤4B)

Name: Abertura
Creator: Artificial Analysis
License: https://artificialanalysis.ai/docs/legal/Terms-of-Use.pdf

Modelos de IA de pesos abertos com até 4B parâmetros. Geralmente, são os menores modelos em termos de demanda de recursos.

Os modelos são considerados de pesos abertos (também chamados frequentemente de código aberto) quando seus pesos estão disponíveis para download. Isso permite a auto-hospedagem em sua própria infraestrutura e a personalização do modelo, por exemplo, por meio de ajuste fino.

Para mais detalhes sobre nossa metodologia, consulte as perguntas frequentes.

K2 Horizon 3.7B e

MiniCPM5-2B são os modelos com maior inteligência entre os modelos muito pequenos de pesos abertos, definidos como aqueles com ≤4B parâmetros, seguidos por

G9v3-3B e

Granite 4.2 3B.

Destaques

Abertura

Artificial Analysis Openness Index · Higher is better

Updated

Inteligência

Artificial Analysis Intelligence Index · Higher is better

Parâmetros totais

Parâmetros treináveis em bilhões

Abertura

Artificial Analysis Openness Index: Score

Openness Index assesses model openness on a 0 to 100 normalized scale (higher is more open)

Inteligência

Artificial Analysis Intelligence Index

Artificial Analysis Intelligence Index v4.3 incorporates 10 evaluations: AA-Briefcase, GDPval-AA v2, AutomationBench-AA, Terminal-Bench v4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, AA-LCR v1.1

Estimate (independent evaluation forthcoming)

Artificial Analysis Intelligence Index v4.3 includes: AA-Briefcase, GDPval-AA v2, AutomationBench-AA, Terminal-Bench v4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, AA-LCR v1.1. See Intelligence Index methodology for further details, including a breakdown of each evaluation and how we run them.

Intelligence Evaluations

Intelligence evaluations measured independently by Artificial Analysis · Higher is better

AA-Briefcase

Agentic knowledge work, (Elo-500)/2000

GDPval-AA v2

Agentic real-world work tasks, (Elo-500)/2000

AutomationBench-AAUpdated

Agentic SaaS workflows

Terminal-Bench v4.0New

Agentic coding & terminal use

SciCode

Coding

Humanity's Last Exam

Reasoning & knowledge

GDP.pdfNew

Professional document reasoning, All-pass

CritPt

Physics reasoning

AA-Omniscience Accuracy

Knowledge

AA-Omniscience Non-Hallucination Rate

1 - hallucination rate

AA-LCR v1.1

Long context reasoning

Harvey LAB-AA

Legal agentic work, criterion pass rate

No data available

EnterpriseOps-Gym-AA

Agentic business operations

No data available

AA-AnalystAgent

Quantitative analysis on spreadsheets & documents

No data available

IFBench

Instruction following

𝜏³-Banking

Agentic tool use

APEX-Agents-AA

Long-horizon agentic tasks

No data available

ITBench-AA

Kubernetes incident root-cause analysis

No data available

MMMU-Pro

Visual reasoning

While model intelligence generally translates across use cases, specific evaluations may be more relevant for certain use cases.

Tamanho

Model Size: Total and Active Parameters

Comparison between total model parameters and parameters active during inference

The total number of trainable weights and biases in the model, expressed in billions. These parameters are learned during training and determine the model's ability to process and generate responses.

The number of parameters actually executed during each inference forward pass, expressed in billions. For Mixture of Experts (MoE) models, a routing mechanism selects a subset of experts per token, resulting in fewer active than total parameters. Dense models use all parameters, so active equals total.

Intelligence Index vs. Active Parameters

Artificial Analysis Intelligence Index · Active parameters at inference time

Most attractive quadrant

Pareto line

Intelligence Index vs. Total Parameters

Artificial Analysis Intelligence Index · Size in parameters (billions)

Most attractive quadrant

Pareto line

The total number of trainable weights and biases in the model, expressed in billions. These parameters are learned during training and determine the model's ability to process and generate responses.

Janela de contexto

Context Window

Context window: tokens limit · Higher is better

Larger context windows are relevant to RAG (Retrieval Augmented Generation) LLM workflows which typically involve reasoning and information retrieval of large amounts of data.

Maximum number of combined input & output tokens. Output tokens commonly have a significantly lower limit (varied by model).

Mais detalhes

							Pesos		Benchmarks de provedores
K2 Horizon 3.7B	MBZUAI Institute of Foundation Models	16	3.7B	524k	-	-	🤗	-	Ver
MiniCPM5-2B	OpenBMB	13	2.6B	131k	-	-	🤗	-	Ver
G9v3-3B	AI9Stars	11	3B	131k	-	-	🤗		Ver
Granite 4.2 3B	IBM	9	3B	131k	$0.0	217	🤗		Ver
MiniCPM5-1B (Reasoning)	OpenBMB	9	1B	128k	-	-	🤗	-	Ver
MiniCPM5-1B (Non-reasoning)	OpenBMB	9	1B	128k	-	-	🤗	-	Ver
Nanbeige4.1-3B	Nanbeige	8	3.9B	256k	-	-	🤗	-	Ver
LFM2.5-2.6B	Liquid AI	8	2.7B	128k	-	-	🤗		Ver
NVIDIA Nemotron 3 Nano 4B	NVIDIA	7	4.0B	262k	-	-	🤗	-	Ver
Qwen3.5 2B (Reasoning)	Alibaba	7	2.3B	262k	-	-	🤗	-	Ver
Phi-4 Mini Instruct	Microsoft	6	3.8B	128k	-	45	🤗		Ver
Qwen3.5 2B (Non-reasoning)	Alibaba	6	2.3B	262k	-	-	🤗	-	Ver

Comparação de modelos muito pequenos de IA de pesos abertos (≤4B)

Abertura

Inteligência

Parâmetros totais

Abertura

Artificial Analysis Openness Index: Score

Inteligência

Artificial Analysis Intelligence Index

Artificial Analysis Intelligence Index

Intelligence Evaluations

Intelligence Evaluation Relevance

Artificial Analysis Intelligence Index

Tamanho

Model Size: Total and Active Parameters

Total Parameters

Active Parameters at Inference Time

Intelligence Index vs. Active Parameters

Artificial Analysis Intelligence Index

Active Parameters at Inference Time

Intelligence Index vs. Total Parameters

Artificial Analysis Intelligence Index

Total Parameters

Janela de contexto

Context Window

Context Window for RAG

Context Window

Mais detalhes