An Empirical Study on Quantized LLMs for Legacy Code Modernization
REMODEL-LLM: Transforming C code to Java using LLMs
This paper investigates the efficacy of 19 small, quantized LLMs (under 20 billion parameters) for the C to Java translation task. We use a novel, hybrid pipeline that leverages Abstract Syntax Trees (ASTs) for semantic decomposition and employs a highly constrained, rule based prompting strategy. The results are stark: a clear multi tiered performance divide emerged.
Executive Impact: Key Findings for Enterprises
Understanding the immediate implications for enterprise software development and strategic AI adoption.
13/20
Max Tests Passed
0%
Tier 3 Success Rate
50%+
Tier 1 Pass Rate
Key Takeaway: Quantized LLMs show potential but face hard ceilings on complex semantic transformations like function pointers and sizeof. Strategic prompting is key, but fundamental paradigm shifts remain challenging.
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Methodology
Performance Tiers
Key Findings
Methodology
Explores the hybrid AST-driven pipeline and guardrail-driven prompting strategy.
Enterprise Process Flow
C Repository
→
C Code Analysis (AST)
→
Global Variables
→
Functions
→
Structures
→
LLM Translation
→
Java Repository
→
Compile and Compare Outputs
Performance Tiers
Details the stark multi-tiered performance divide among the 19 models tested.
| Tier | Description | Key Characteristics | Models |
|---|---|---|---|
| Tier 1 (Viable) | Passed >50% of tests. Handled complex rules, but failed on most advanced concepts. |
|
deepseek-coder-v2, codeqwen, phi4 |
| Tier 2 (Flawed but Occasionally Successful) | Passed some tests (20-40%). Showed flawed semantic understanding and copied C-like syntax. |
|
mistral-nemo, mistral |
| Tier 3 (Complete Failure) | 0% success rate. Unable to generate basic runnable Java boilerplate. |
|
llama3.1, gemma3, starcoder2, etc. |
Key Findings
Highlights specific successes and universal failures.
1/19
Model passed T3 (Union) and T8 (Bitfields) due to superior rule application.
Case Study: Pointer Arithmetic (T1)
Highlight: deepseek-coder-v2's successful refactoring of C pointers to Java array indices.
Test Case 1 (T1) was a foundational test for C pointer arithmetic. The deepseek-coder-v2 model demonstrated perfect adherence to prompt rules, refactoring C pointer logic into Java array index logic using a separate index variable.
Case Study: goto Statement (T13) Failure
Highlight: mistral-nemo's syntactically invalid translation of C goto to Java 'continue'.
Test Case 13 (T13) presented a harder C specific control flow problem, requiring refactoring a goto-based loop. mistral-nemo produced a syntactically invalid translation by naively mapping 'goto' to 'continue', demonstrating its inability to perform semantic restructuring.
Calculate Your Potential ROI
Estimate the potential return on investment for modernizing legacy C code to Java with AI assistance.
Your AI-Powered Modernization Roadmap
A phased approach to integrate AI-powered code translation into your enterprise workflow.
Phase 1: Pilot & Assessment
Translate a small, critical C module to Java, assess translation quality, and refine AI prompting strategies.
Phase 2: Tooling Integration
Integrate REMODEL-LLM pipeline with existing CI/CD, establish automated verification, and build internal expertise.
Phase 3: Iterative Modernization
Expand to larger C codebases, focusing on modules with high maintenance burden or security risks, with human-in-the-loop review.
Phase 4: Ecosystem & Optimization
Leverage Java ecosystem benefits, optimize translated code for performance, and continuous monitoring.
Ready to Transform Your Legacy Code?
Connect with our experts to discuss how REMODEL-LLM can accelerate your C to Java migration.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat