BlueLM-2.5-3B: Compact Multimodal Edge LLM

• BlueLM-2.5-3B is a compact, 2.9B-parameter multimodal large language model designed for edge devices, supporting both rapid direct responses and extended chain-of-thought reasoning.
• It integrates a tiled ViT visual encoder, a two-layer adapter MLP, and a dense LLM backbone to achieve robust and efficient alignment of text and vision inputs.
• The model features dual 'thinking' and 'non-thinking' modes with explicit token budget control, enabling precise trade-offs between computational efficiency and response quality.

BlueLM-2.5-3B is a compact, unified dense multimodal LLM (MLLM) engineered for efficient deployment on edge devices such as smartphones, AR glasses, and mobile robots. With a total parameter count of 2.9 billion (0.4B in the vision encoder; 2.5B in the LLM core), it aims to deliver strong general-purpose and high-level reasoning capabilities in both single-modal (text) and multimodal (vision+text) tasks, offering fine-grained control over computational cost and inference style. BlueLM-2.5-3B distinguishes itself as the first 3B-scale MLLM to support both "thinking" (multi-step chain-of-thought) and "non-thinking" (fast, direct response) inference modes, with explicit token budget controls leveraged through a control token │BlueThink│.

1. Motivation and Rationale

BlueLM-2.5-3B was motivated by the growing need for high-level multimodal reasoning on resource-constrained edge devices. Conventional cloud-hosted MLLMs present challenges of latency, increased energy consumption, and privacy risks. By targeting edge-side execution, BlueLM-2.5-3B introduces a deployment paradigm that emphasizes:

• Compactness: At 2.9B parameters, the model is 22% smaller than Qwen2.5-VL-3B (3.76B) while maintaining competitive capacity.
• Efficiency: The visual encoder employs tiled ViT architectures supporting image resolutions up to 1512×1512 with near-linear scaling in throughput.
• Inference Flexibility: Via [│BlueThink│], the model can explicitly choose between low-latency, brief-answer "non-thinking" mode (≤4K output tokens), and extended chain-of-thought "thinking" mode (up to 32K tokens context), with robust, prompt-level control and failure rate under 1 PPM (Xiong et al., 8 Jul 2025).

This dual-mode design enables on-device AI systems to dynamically manage power, latency, and answer quality as demanded by the application context.

2. Model Architecture

BlueLM-2.5-3B comprises three principal components, jointly enabling dense multimodal processing:

Component	Parameter Count	Key Design Features
ViT Visual Encoder	0.4B	27 Transformer layers; SigLIP2 “so400m-patch14-384”
Adapter MLP	–	2-layer MLP; 2×2 spatial downsampling
Dense LLM Backbone	2.5B	Transformer; details protected; optimized multimodal

• ViT Visual Encoder: Initialized from SigLIP2, with AnyRes tile-based processing (default: 4×4 grid of 378×378 patches), enabling high-throughput vision input.
• Adapter MLP: Two-layer projection from ViT token space to the LLM embedding space, including spatial downsampling for memory/thrift alignment.
• LLM Backbone: Proprietary in architectural detail but accounts for the majority of the model’s parameter footprint, optimized for dense multimodal alignment (Xiong et al., 8 Jul 2025).

Notably, BlueLM-2.5-3B's parameterization is substantially more compact than leading open-source models such as Qwen2.5-VL-3B and Kimi-VL-A3B-16B.

3. Training Methodology and Data Pipeline

BlueLM-2.5-3B's data curation and training pipeline are designed for data efficiency and capability coverage:

Data Sources and Mixture

• Pure-text (9.3T tokens): Multilingual corpora, code, Q&A, academic papers, exams.
• Image–text pairs (2.4T): Captions, OCR, VQA, GUI data.
• Synthetic reasoning data (3.3T): Chain-of-thought (CoT) for STEM, multimodal VQA.
• Long-context sequences: Text and interleaved text-image data for 32K context support.

Pipeline Innovations

• PreSelect: Capability-driven sample selection.
• DSIR: Long-tail upsampling.
• RegMix: Dynamic multimodal/text ratio, finalized at 60% image–text : 40% pure-text.
• Multi-sample concatenation: Packs semantically similar samples for optimal GPU throughput.

Staged Training

• a) LLM pre-training: 9.3T tokens distilled from a 7B teacher; exported to LLM-base-3B.
• b) Joint multimodal pre-training: 4T tokens, 60:40 image–text to text, 4,096 sequence length.
• c) Reasoning-enhanced pre-training: 2.5T tokens with long-chain CoT up to 4,096 tokens.
• d) Long-context adaptation: 1.3T tokens, RoPE→YaRN transition, context up to 32K tokens (exceeding standard MLLMs).

Supervised Fine-Tuning and RLHF

• Supervised fine-tuning: Joint multimodal/text-instruction data, [│BlueThink│] mode switch, and 300K high-quality STEM CoT samples with answer verification.
• Hybrid heterogeneous RLHF: Combines open/preference data, multiple reward models (discriminative/generative), explicit rule-based penalties (correctness, format, repetition), and a dynamic length-penalty to discourage inefficient overthinking:

$$L_{\min}=\min_i L_i,\; L_{\max}=\max_i L_i,\; \Delta_L = \max(500, L_{\max}-L_{\min})$$

$$R_{\text{len}(i)} = \alpha\,\lambda_i,\quad\lambda_i=0.5-\frac{L_i-L_{\min}}{\Delta_L}$$

$$\mathcal{L}_\text{RL} = \mathbb{E}\left[A_i\log\frac{\pi_\theta(y_i|x)}{\pi_\mathrm{ref}(y_i|x)}-\beta\,\mathrm{KL}(\pi_\theta\|\pi_\mathrm{ref})\right]$$

where $\alpha$ decays over training but is floored at 0.2.

• Infrastructure: Scaled on 2,048 GPUs with 4×200 Gb/s InfiniBand, context/data parallelism in ViT, asynchronous RL pipeline, and load-balanced vLLM inference (Xiong et al., 8 Jul 2025).

4. Performance, Benchmarks, and Data Efficiency

BlueLM-2.5-3B demonstrates robust performance on comprehensive evaluations across multimodal and text-only tasks, using temperature=0.6, top-p=0.95, and top-k=20.

Thinking Mode (32K context, CoT)

• Multimodal benchmarks: Trails only Kimi-VL-A3B-16B by ~5% on average; outperforms Qwen2.5-VL-72B on MathVision/MathVista.
• Text-only benchmarks: Matches Qwen3-4B performance (e.g., GSM8K: 95.5% vs. 94.7%), and surpasses Qwen2.5-VL-72B on Math-500/GSM8K.

Non-Thinking Mode (4K output, direct answer)

• Multimodal reasoning: Exceeds Qwen2.5-VL-3B by 5–10 points on MathVista (70.8 vs. 60.1) and MathVision (28.5 vs. 21.2). Outperforms Gemma-3-4B on 9/10 tasks; within 5% of Kimi-VL-A3B on 6/10.
• Text-only: Outpaces Qwen2.5-VL-3B on Math-500 (80.0 vs. 58.0) and BBH (80.6 vs. 45.7), leading Gemma-3-4B on 9/11 tasks.
• GUI grounding: ScreenSpot vivo: 89.8%, matches UI-TARS-2B, narrows the gap to UI-TARS-7B.

Data Efficiency

• Pre-training footprint: 17.1T tokens vs. Qwen2.5-VL-3B’s 22.1T (–23%) and Qwen3-4B’s 36T (–53%).
• Benchmark parity/lead: Consistent competitive or superior results despite reduced data (Xiong et al., 8 Jul 2025).

5. Inference Modes and Token Budget Control

BlueLM-2.5-3B formalizes the distinction between two inference styles:

• Non-thinking mode: Yields short, rationale-free generation with a strict 4K token limit, minimizing computational cost for trivial queries.
• Thinking mode: Activated by appending [│BlueThink│] to the prompt; enables multi-step reasoning, leveraging up to 32K context window. This switch is explicit and designed to robustly activate reasoning mode (failure rate < 1 PPM), and the application can control the precise token budget for reasoning, providing granular trade-offs between answer quality and resource usage.

This mechanism allows edge applications to dynamically adapt to user and task requirements, optimizing for either latency or depth of output.

6. Practical Implications and Limitations

BlueLM-2.5-3B demonstrates that sub-3B parameter scale MLLMs can match or approach the performance of much larger cloud-scale models in both dense multimodal understanding and high-level reasoning, while enabling on-device inference with constrained hardware.

• Edge deployment: Token budget switching allows real-time adaptation between resource usage and reasoning quality, critical for user-facing mobile applications.
• Current constraints: The input space is limited to visual+text modalities (audio/speech unsupported), and LLM hyperparameters remain proprietary, restricting full open-source re-implementation.
• Planned extensions: Future work involves tri-modal (text, vision, speech) support, richer synthetic reasoning data for generalization (e.g., cross-lingual CoT), and dynamic chain-of-thought stopping when marginal utility is low (Xiong et al., 8 Jul 2025).

7. Context and Related Work

BlueLM-2.5-3B extends the roadmap of mobile-friendly MLLMs, contrasting with contemporaries such as BlueLM-V-3B (Lu et al., 2024), which focuses on algorithmic and system co-design (relaxed dynamic image resolution, token downsampling, mixed-precision quantization) for real-time on-device inference. While both models share similar scale (2.9B vs. 3B parameters), BlueLM-2.5-3B’s explicit dual-mode reasoning and advanced data/optimization pipeline differentiate its scientific and practical profile within the edge AI community.

In sum, BlueLM-2.5-3B represents a paradigm shift toward robust, scalable, and user-adaptive multimodal intelligence on edge platforms. Its dual-mode inference, rigorous data management, and explicit reasoning controls mark an important inflection point for high-performance MLLMs under severe power and memory budgets (Xiong et al., 8 Jul 2025).