Robots Atlas>ROBOTS ATLAS
ERNIE 4.5

ERNIE 4.5

ERNIE 4.5 (do VL-424B-A47B) · Family: ERNIE
Baidu’s open-source (Apache 2.0) multimodal foundation model: heterogeneous MoE, variants up to 424B-A47B (VL), 128K context; powers the Ernie Bot assistant.
✓ Active✓ Public access⚖ Open sourceLLMMultimodalVision📁 ERNIE
Context window
128K
tokens
Parameters
do 424B (47B aktywne) — VL-424B-A47B; warianty 300B-A47B, 21B-A3B, 28B-A3B, 0.3B
parameters
Release date
16 March 2025
Access:DownloadAPIHostedDeployment:💻 Local☁ Cloud

Overview

ERNIE 4.5 is a multimodal foundation model developed by Baidu, announced on 16 March 2025 alongside the ERNIE X1 reasoning model. It belongs to the ERNIE family (Chinese 文心, Wenxin) that powers the Ernie Bot assistant.

The model is built on a heterogeneous multimodal mixture-of-experts (MoE) architecture that shares parameters across modalities while also allocating dedicated parameters for each. The language (LLM) versions handle text, while the vision-language (VL) versions additionally support image and video; the context window is 128K tokens.

Baidu released the ERNIE 4.5 family as open source under the Apache 2.0 license, publishing weights and inference code on Hugging Face and AI Studio. The open variants include the ERNIE-4.5-300B-A47B language model (300B total / 47B activated), 21B-A3B (21B / 3B) and a dense 0.3B, as well as the vision-language models ERNIE-4.5-VL-424B-A47B (424B total / 47B activated) and VL-28B-A3B.

In April 2025, at the Baidu Create 2025 conference, an optimized ERNIE 4.5 Turbo variant (faster responses, lower cost) was also released. The models are available through the Ernie Bot assistant (yiyan.baidu.com, offered free of charge), the Baidu AI Cloud (Qianfan) platform, and locally thanks to the open weights.

Classification
LLMMultimodalVision
Family: ERNIE
Access & deployment
DownloadAPIHosted
LocalCloud
Weights: Open source
Key parameters
📏 Context: 128K
🧩 Parameters: do 424B (47B aktywne) — VL-424B-A47B; warianty 300B-A47B, 21B-A3B, 28B-A3B, 0.3B
Tools · ✓ Fine-tuning
📥 Input: text, image, video

Technical specification

Context window
128K
tokens
Parameters
do 424B (47B aktywne) — VL-424B-A47B; warianty 300B-A47B, 21B-A3B, 28B-A3B, 0.3B
parameters
License
Apache 2.0
Hardware requirements
An open-weights family (Apache 2.0) on Hugging Face and AI Studio. The largest MoE variants (up to 424B total parameters) require a large-VRAM GPU cluster; smaller variants (21B-A3B, 0.3B) and quantized versions (FP8, W4A8C8) enable lighter deployments. Also available via Baidu Cloud (Qianfan).
Features:Tool useFine-tuning
Modalities
⬇ Input
textimagevideo
⬆ Output
text

Capabilities and applications

Native model capabilities
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Mathematical reasoning
The model's ability to solve mathematical tasks requiring multi-step reasoning — equations, proofs, combinatorics, geometry, calculus and competition-level problems.
Category: reasoning
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Long context
Support for large context windows — tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Image understanding
Analysing and interpreting the content of images.
Category: vision
Video understanding
The model's ability to analyse and interpret video content — recognising actions, motion, events and relationships between objects over time.
Category: video
Multimodal understanding
Category: multimodal
Tool use
The model's ability to call external functions, APIs and tools during a conversation: calculator, search engine, code editor, database. The model decides when and how to use a tool and interprets its result.
Category: planning
Instruction following
Precisely following instructions contained in the prompt: response format, length, style, constraints (e.g. 'reply in six words'). GPT-5.1 significantly improved this capability compared to GPT-5.
Category: language
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Natural conversation
Conducting a conversation with a tone close to human: a warmer voice, empathy in emotional responses, humour, and avoiding the stiff 'AI assistant jargon'. Introduced as a deliberate improvement in GPT-5.1 Instant.
Category: language

Technical architecture

Core Architecture
Training Techniques