Decoding LLM Upgrades in Finance: From General Capabilities to Professional Workflows

1. Background: Why Finance is a High-Value Scenario for LLMs

The financial industry is inherently characterized by data density, complex documentation, stringent compliance requirements, and extended decision-making chains. Traditionally, institutions rely heavily on manual labor to process unstructured information across investment research, risk management, compliance, and operations—dealing with earnings reports, announcements, research papers, contracts, and regulatory filings. Large Language Models (LLMs) offer a new technical pathway for automating and enhancing these processes.

However, financial scenarios demand far greater accuracy, traceability, and long-document comprehension than typical conversational use cases. A general-purpose chatbot might excel at common sense Q&A but often fails to deliver reliable results for tasks like “cross-referencing financial data across documents” or “deducing risk exposure from portfolio changes.” Therefore, building LLM capabilities for finance focuses not merely on scaling parameter sizes, but on domain reinforcement during the post-training phase, toolchain integration, and workflow orchestration.

A recently released LLM version follows this exact trajectory. While maintaining a stable base parameter scale, the primary improvements stem from increased compute allocation during post-training and adjustments to the training framework. This has resulted in significant performance gains in specialized evaluations like financial research and modeling. This article will analyze these upgrades across four dimensions: technical capabilities, evaluation performance, industry solutions, and implementation practices.

2. Core Upgrade Path: Stable Base, Post-Training Focus

2.1 Parameter Scale and Context Window

The new version retains the foundational architecture of its predecessor, featuring a total parameter scale at the trillion level, activated parameters at the billion level, and a context window supporting millions of tokens. This configuration provides two main benefits:

• Activated parameters control inference costs: By utilizing sparse activation under a sufficiently large total capacity, the computational overhead per inference is reduced, keeping unit costs manageable in high-concurrency scenarios.
• Million-level context supports long-document tasks: In finance, a single prospectus, annual report, or a set of multi-period financial statements often exceeds the context limits of standard models. Long context capabilities allow the model to correlate information across different chapters and documents in a single inference pass.

2.2 Compute Expansion in Post-Training

The core incremental improvement in this upgrade comes from post-training. The R&D team increased compute resources for the reinforcement learning (RL) phase and adjusted the training framework, long-context inference, sandbox infrastructure, and scheduling system to support joint training across hundreds of downstream domains. The engineering significance of this includes:

• Multi-domain joint training: Capabilities across finance, coding, 3D creation, and long-horizon agentic tasks can share underlying representations while receiving targeted optimization for their specific task distributions.
• Combining RL with tool usage: Financial tasks frequently require calling external data sources, executing calculations, and verifying documents. Introducing reinforcement signals for tool calling and multi-step execution during post-training makes the model’s performance more stable in real-world workflows.

2.3 Capability Coverage

Beyond finance, this upgrade covers code generation, 3D creation, and long-horizon agentic tasks. These capabilities exhibit synergies with financial scenarios:

• Coding capabilities are utilized for formula implementation and data processing script generation in financial modeling.
• 3D creation corresponds to general multimodal capabilities; while not directly serving finance, it reflects the base model’s overall generalization level.
• Long-horizon agentic tasks align with multi-step operations in investment research workflows, such as “screening targets -> pulling financial data -> calculating metrics -> generating reports.”

3. Evaluation Performance: Significant Gains in Financial Dimensions

3.1 Financial Document Comprehension and Numerical Reasoning

In an evaluation examining complex financial data analysis from real-world documents, the new version scored 64.4. This evaluation involves information retrieval, cross-document verification, and multi-step numerical reasoning. Typical tasks include:

• Extracting the same metric from multiple financial reports and checking for consistency;
• Recalculating ratios after adjusting the scope based on footnote information;
• Determining trends by combining management discussion with financial data.

Such tasks demand extremely high long-context localization capability and numerical reasoning accuracy, representing the core challenges that distinguish financial scenarios from general Q&A.

3.2 Complete Investment Workflow

In an evaluation measuring a complete investment workflow, the new version showed improvement over its predecessor. The testing covered the entire process from target screening to profitability analysis. This indicates that the model can maintain logical coherence not just for isolated questions, but within multi-step, stateful workflows.

3.3 Internal Specialized Financial Tests

Officially disclosed internal test data shows:

Test Dimension	Previous Score	New Version Score	Improvement
Financial Research	58.8	70.4	+11.6
Financial Modeling	50.4	57.6	+7.2

The over 11-point increase in financial research indicates a qualitative improvement in the model’s ability regarding information integration, viewpoint extraction, and report generation. The 7.2-point increase in financial modeling reflects progress in structured calculation, hypothesis testing, and sensitivity analysis.

3.4 General Capability Comparison

In an agentic terminal task evaluation, the new version achieved 65.7%, higher than a mainstream model’s 58.2% and slightly below another model’s 66.4%. This test evaluates a coding agent’s tool usage and multi-step execution capabilities in a real-world, resource-constrained terminal environment. In an independent evaluation agency’s agentic index, the new version scored 55, ranking fourth among tracked models, an 11-point increase from the previous version.

In a tool-assisted CAD test, the new version scored 92.7, showing a gap compared to top-tier models (96.2 and 95.9). This suggests that general models still have room for improvement in highly specialized 3D/engineering toolchains. However, while coding and 3D tests reflect general capabilities, the financial document and investment workflow tests are much closer to the actual needs of institutional business operations.

4. Financial Industry Solution: Data Sources, Skills, and Compliance

4.1 Solution Architecture

Prior to the new model release, the team launched a financial industry solution comprising core components:

• Data source integration: Connected to over a dozen authoritative data sources covering market quotes, financial reports, announcements, macroeconomics, and industry categories.
• Built-in financial skills: Includes 9 financial skills such as financial metric calculation, valuation modeling, risk metric measurement, announcement summarization, and research report generation.
• Compliance and security measures: Built-in 5 compliance and security measures, including data masking, access control, output review, audit logs, and permission isolation.

4.2 Covered Scenarios

The solution covers the following typical scenarios:

1. Portfolio Morning Briefing: Automatically aggregates portfolio changes, anomaly alerts, and relevant announcements to generate a morning report.
2. Earnings Report Commentary: Parses key subjects of earnings reports, compares them with market expectations, and outputs commentary highlights.
3. Project Screening: Screens candidates from a project repository based on investment strategies and outputs an initial screening report.
4. In-Depth Research: Integrates multi-source information around specific targets or industries to generate in-depth research reports.
5. Portfolio Review: Analyzes portfolio return attribution, risk exposure, and rebalancing effectiveness.

A common characteristic of these scenarios is: multi-source input, multi-step process, and the need for explainable output. The LLM acts not as a “human replacement,” but as an accelerator for “information processing and draft generation.”

4.3 Partnership Models

According to disclosures, dozens of financial institutions have engaged in business innovation based on this model, its enterprise version, and workflow products, or have co-built financial solutions with the team. These institutions include brokerages, commercial banks, venture capital funds, and investment research data agencies. Collaboration formats include:

• Model API Calls: Integrating model capabilities into existing systems via APIs.
• Enterprise Edition Deployment: Deploying the enterprise version in private or dedicated environments to meet data residency requirements.
• Joint Development: Co-designing financial skills, workflows, and compliance strategies with the institutions.

The head of the team’s enterprise business previously stated to media that the company does not focus on entertainment scenarios, but rather on “hard but correct” productive domains like coding, finance, legal, and scientific research. This strategic choice dictates the design orientation of its financial solution: workflow-heavy, compliance-heavy, and integration-heavy, rather than pursuing general conversational entertainment.

5. Developer Perspective: Designing Financial AI Workflows

For developers and enterprise users looking to implement LLM capabilities in financial scenarios, the following layers are typically required:

5.1 Data Layer

• Structured data: Market quotes, financial data, and portfolio data accessed via databases or APIs.
• Unstructured data: Announcements, research reports, contracts, and regulatory files processed via document parsing and vectorized storage.
• Data governance: Version management, access control, and desensitization processing.

5.2 Model Layer

• Base selection: Choosing the appropriate parameter scale and context window based on task complexity.
• Post-training adaptation: Conducting domain fine-tuning or reinforcement learning on a general base model to inject financial knowledge and workflow preferences.
• Tool calling: Enabling the model to call calculators, database queries, document retrieval, and other tools via function calling or agentic frameworks.

5.3 Orchestration Layer

• Workflow engine: Defining the states and transition conditions of multi-step tasks.
• Sandbox environment: Providing an isolated environment for code execution and data calculation.
• Scheduling system: Managing concurrent requests, timeout retries, and priority queues.

5.4 Compliance Layer

• Input filtering: Intercepting sensitive information.
• Output review: Conducting compliance checks on generated content.
• Audit logs: Recording every model call and human intervention.

A simplified pseudo-code example of a financial workflow:

def financial_research_workflow(query):
    # 1. Parse user intent
    intent = model.parse_intent(query)
    # 2. Retrieve relevant documents
    docs = retrieve_documents(intent.entities, intent.time_range)
    # 3. Extract key data
    data = model.extract_financial_data(docs)
    # 4. Execute calculations
    metrics = sandbox.run(calculate_metrics, data)
    # 5. Generate report
    report = model.generate_report(intent, data, metrics)
    # 6. Compliance review
    report = compliance_review(report)
    return report

6. Conclusion and Outlook

The core logic of this LLM upgrade can be summarized as: stable base, post-training focus, and financial breakthroughs. By increasing RL compute and adjusting the training framework and scheduling system, the model has achieved measurable improvements in financial research, financial modeling, long-context reasoning, and agentic tasks. The accompanying financial industry solution bridges the gap between “model capability” and “business implementation” across three dimensions: data sources, skills, and compliance.

Key trends worth noting for developers and enterprise users include:

1. The competitive focus of financial AI is shifting from “general capabilities” to “workflow capabilities”: Those who can better orchestrate multi-step tasks, integrate tools, and satisfy compliance requirements will establish themselves in institutional scenarios.
2. Post-training and domain adaptation become key differentiators: As base parameter scales stabilize, injecting domain knowledge and workflow preferences becomes the primary means of enhancing professional performance.
3. Compliance and security are prerequisites for financial implementation: A solution lacking a compliance layer will struggle to enter the production environment of financial institutions.

According to the official timeline, full model weights will be publicly available soon. At that point, developers can evaluate its performance on specific financial tasks within their own environments and design workflows tailored to their business needs. The implementation of financial AI is transitioning from “usable” to “effective” and “controllable.”