BlueHeart large model BlueLM

What is the BlueHeart large model, BlueLM?

BlueLM, the large-scale model from BlueHeart, is a universal large-model series developed independently by vivo’s Global AI Research Institute. Its capabilities include content creation, knowledge answering, logical reasoning, code generation, and information extraction. It serves both vivo’s operating systems and various applications, and it also makes some of the model weights and deployment codes available to developers.

It is important to distinguish between the overall product portfolio of BlueHeart large models and the open-source BlueLM-7B project. The versions that can be downloaded publicly are the 7B base model, the dialogue model, the 32K version, and the quantized model; this does not mean that all of vivo’s models for edge devices, cloud environments, image processing, speech processing, or multi-modal tasks are available in open source form.

Available model versions

Base models are more suitable for research, extension, and fine-tuning, while dialogue models have been optimized for interaction. 32K indicates the ability to handle longer contexts, but the actual usable length, speed, and memory usage are influenced by the framework, batch size, precision, and hardware.

Main functions

Content creation and rewriting

• Dialogue models can generate articles, summaries, emails, marketing drafts, and structured content based on instructions in Chinese or English.
• Users can specify the topic, tone, length, target audience, and formatting requirements; after the model generates the text, fact-checking and language review are carried out.
• The base model is suitable for text continuation and domain fine-tuning, while the dialogue model is better suited for direct instruction-based interactions; choosing the wrong one can affect the user experience.

Knowledge Q&A and information extraction

• The model can answer questions related to common knowledge, work skills, and information contained in texts; it can also extract entities, key points, fields, or relationships from given materials.
• The 32K version is suitable for summarizing, answering questions, or extracting information from longer documents within a given context, but it does not automatically connect to the latest databases.
• Model knowledge may be outdated or lead to errors; regulations, medical, financial, news, and corporate data must be retrieved from reliable sources and verified manually.

Logical reasoning and code generation

• BlueLM is capable of performing reasoning related to common sense, science, mathematics, and task planning; it can also generate and interpret code.
• Developers should provide clear constraints, examples, and test cases, and run the generated code in an isolated environment to avoid executing unknown commands directly.
• The publicly available benchmark scores reflect only specific datasets and evaluation settings; they do not mean that the same level of accuracy can be achieved in all real-world applications.

Long text processing

• Base-32K and Chat-32K are used for understanding longer contexts, enabling the processing of long reports, multiple sections of text, extended conversations, as well as document-based question answering.
• Long texts increase memory usage, inference latency, and computational costs; it is necessary to test the length of the target documents as well as the level of concurrency before deployment.
• Placing the content within a 32K context does not mean that the model will remember every detail; important fields still need to be extracted and verified in a structured manner.

Tool invocation and local APIs

• BlueLM-7B-Chat-32K supports function calling, allowing natural language requests to be converted into application-defined functions and parameters.
• The official repository provides local API examples in the OpenAI format, supporting streaming responses and tool calls, which facilitates integration with existing chat application backends.
• The model is responsible only for specifying the intention of the call; actual permissions for the tool, parameter validation, execution, retry attempts, and result auditing must be handled by the application.
• This API example runs on the user’s own server; it is not hosted on Vivo’s public cloud services, and there is no official pay-as-you-go pricing.

Training language and input/output

Local Deployment Tutorial

1. First, read the Open Atom Model License to determine whether research, internal use, distribution, or commercial delivery are in compliance with its terms.
2. Select the Base, Chat, 32K, or quantized model based on the task, and assess the compatibility of GPU memory, disk storage, CUDA, drivers, and the inference framework.
3. Obtain the official repository and the corresponding model files, install the Python dependencies specified in the project; when loading the model, examine the remote code that needs to be executed.
4. First, run the command-line or web-based demo to test the performance with Chinese text, long texts, code, refusal responses, and content from various business domains, using a fixed test set.
5. When serviceization is required, start the local API example or deploy vLLM, and add authentication, rate limiting, timeouts, log masking, and monitoring.
6. Before integrating the service, steps such as search, tool whitelist, output filtering, and manual review are carried out, followed by tests for concurrency, long texts, and fault recovery.
7. Record the model version, quantization method, prompt, license, and data acquisition method to facilitate upgrades, audits, and issue rollback.

Quantization and hardware selection

The 5GB of video memory is a reference value based on an official 4-bit example; it should not be applied to all systems, batch sizes, or context lengths. Developers must verify the compatibility and performance of CPUs, Apple chips, and other accelerators on their own.

Fine-tuning and secondary development

• The warehouse provides a comprehensive supervision fine-tuning process, training scripts, and sample data, which can be used for domain-specific question answering, style adaptation, or learning task formats.
• Before fine-tuning, duplicate, erroneous, sensitive, and unauthorized data should be removed, and a separate validation set should be created to prevent the model from memorizing only the training samples.
• Full fine-tuning requires high resource consumption; developers can choose efficient methods by considering the specific task, but they must themselves verify the implementation and its outcomes.
• When redistributing models, code, or services, the statements required by the license must be retained, and the responsibilities regarding data, branding, and the generated content must be reassessed.

Price and usage costs

The fact that model files can be downloaded for free does not mean that there are no costs associated with deployment, nor does it imply that access to Vivo’s public cloud computing resources is guaranteed. There are no issues related to refunds for open-source downloads, but refunds for cloud servers, storage, and third-party model hosting must be handled in accordance with the rules set by the respective service providers.

Differences between APIs, SDKs, and platforms

Open-source licenses and commercial use

• In December 2024, the official BlueLM repository updated the model license to the Open Atom Model License.
• The warehouse is intended for use by developers in academic research and commercial applications; however, any commercial use must still comply with all the conditions of the license as well as relevant laws.
• The license applies only to the models and materials explicitly provided by the warehouse; it does not automatically cover the vivo trademark, consumer applications, cloud services, or other models that are not available.
• When using third-party data, fine-tuning datasets, inference frameworks, and dependencies, it is also necessary to comply with their respective licenses and rules regarding personal information.
• Before preparing to redistribute the model, provide hosting services, or embed it in commercial devices, legal officers should review the statements, ownership provisions, restrictions, and liability terms.

Notes on privacy, security, and copyright

• Local deployment keeps prompts and documents within the user’s controlled environment, but logs, monitoring data, object storage, and external tools may still transmit data.
• Downloading from the model hosting site and using the services within vivo apps are each governed by the account policies of their respective platforms, as well as by rules regarding cookies, logging, and privacy.
• Before loading a model with custom remote code, it is necessary to review the code, fix its version, and isolate the execution environment in order to reduce supply chain risks.
• External APIs require authentication, rate limiting, limits on input length, masking of sensitive information, isolation of dangerous tools, and blocking of abnormal outputs.
• Models may generate factually incorrect information, biases, harmful suggestions, similar texts, or unsafe code; therefore, they must be manually reviewed before being published and executed.
• User inputs, fine-tuning data, and generated results may involve copyright, trade secrets, and personal information; therefore, the deployment party needs to establish procedures for deletion and handling complaints.

Suitable for users and scenarios

• Universities and research teams: conducting experiments, assessing Chinese language skills, and researching long texts as well as tool invocation.
• Enterprise developers: Build prototypes for knowledge questioning, text extraction, writing, or code assistance in their own environments.
• Teams sensitive to privacy: They control the data flow through local deployment, but it is still necessary to audit all stored data as well as any external dependencies.
• Individuals with limited hardware resources: Use 4bits or other quantized versions to experience the 7B dialogue model.
• vivo Ecosystem Developers: Evaluate advanced AI APIs for on-device use, but design them in accordance with the system version and compatible device models.

Advantages and limitations

Main advantages

• Available in public base, dialogue, 32K, and various quantization versions, facilitating gradual verification from research to local application.
• It provides examples of APIs in the forms of reasoning, fine-tuning, vLLM, and OpenAI, thereby reducing the effort required for setting up the underlying infrastructure.
• Chinese and English are the primary languages used for training, making them suitable for text tasks centered on Chinese.

Usage restrictions

• The main focus of the open-source repository is the 7B text model; not all of the capabilities of the BlueHeart large model can be considered as downloadable weights.
• Publishing the model update time, framework dependencies, and hardware requirements may fall behind the latest developments in the ecosystem; therefore, it is necessary to use a compatible version when deploying.
• The 32K context and function calling do not automatically ensure the accuracy of long texts or the security of tools; comprehensive engineering practices are still required on the application side.
• The technical report for BlueLM-2.5-3B does not mean that the model weights are available; therefore, no commitments should be made regarding downloading or local deployment based on this document.

Frequently Asked Questions

Is the BlueHeart large model BlueLM open-source?

Partially open source. The official team has made available the basic version, the dialogue version, the 32K version, and the quantized version of BlueLM-7B, but not all of vivo’s BlueHeart model portfolio, consumer applications, and system capabilities have been made open source.

Can the BlueLM model be used commercially for free?

The model files can be downloaded for free; the repository welcomes commercial use, provided that the Open Atom Model License and relevant laws are complied with. The costs associated with computing resources, data licensing, maintenance, and compliance are the responsibility of the user.

How long of a context does BlueLM support?

The standard 7B version supports a context length of 2K, while Base-32K and Chat-32K allow for a context length of 32K. A longer context requires more memory and increases latency; moreover, it does not guarantee that every detail can be accurately captured.

Does BlueLM offer an online API?

The warehouse provides examples of API in the OpenAI format that can be deployed independently; these are not Vivo’s public cloud APIs. Vivo’s open platform offers additional edge AI capabilities, and the requirements for accessing them as well as the associated costs must be checked in the relevant documentation.

How much video memory is required at a minimum for BlueLM?

The official 4bits dialogue model example requires approximately 5GB of video memory. Other versions are affected by factors such as precision, context, batch size, and the framework used; actual requirements for production deployment should be determined through testing.

Does BlueLM support image input?

The current official open-source repository for BlueLM-7B is a text model; therefore, the results of subsequent multi-modal research cannot be directly used as available image weights. For image-related tasks, it is necessary to check whether the specific model has been made public.

Is Lanxin Qianxun the same as the BlueLM open-source model?

They are different. BlueHeart QianXun is a consumer application, while the BlueLM open-source project provides models and code that developers can download and deploy; their functions, data policies, versions, and licensing terms are all independent of each other.

Summary

The BlueLM large model is suitable for developers who need capabilities in handling Chinese language, working with long texts, enabling quantitative deployment, and accessing local APIs. When making a choice, it is necessary to distinguish between the entire BlueLM series and the publicly available 7B series; then, the appropriate version should be selected based on the context, available memory, and specific tasks. Before using it in commercial applications, it is essential to conduct reviews regarding licenses, data usage, output results, and security aspects.