Luoberta/Abacus-cve

TEXT GENERATIONPricing:Input $0.408 / Cached $0.0816 / Output $1.972Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Feb 1, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Luoberta/Abacus-cve is a 32 billion parameter code model, based on Qwen3-32B, specifically fine-tuned for security vulnerability fixing tasks. It was trained on 4,078 distilled agent traces from CVE reproduction tasks, demonstrating significant improvements across security benchmarks like LiveCVEBench, PatchEval, and Terminal-Bench. This model excels at identifying and patching security vulnerabilities, outperforming several larger code models and approaching the performance of Claude Sonnet 4.5 on these specialized tasks.

Loading preview...

Abacus-cve: Specialized Model for Security Vulnerability Fixing

Abacus-cve is a 32 billion parameter code model developed by Luoberta, built upon the Qwen3-32B architecture. Its primary distinction lies in its fine-tuning for security vulnerability fixing tasks, utilizing 4,078 distilled agent traces derived from approximately 900 CVE reproduction tasks. These traces were generated using Claude Opus 4.5 within a Mini SWE-Agent harness via the CVE-Factory pipeline.

Key Capabilities

  • Enhanced Security Vulnerability Remediation: Demonstrates substantial improvements in fixing security vulnerabilities, as evidenced by its performance on specialized benchmarks.
  • Benchmark Outperformance: Achieves a ~6.8x improvement on LiveCVEBench (from 5.29% to 35.79%), ~4.2x on PatchEval, and ~2.3x on Terminal-Bench compared to its base model. It surpasses Qwen3-Coder-480B, MiniMax-M2, and Claude Sonnet 4 on these security metrics.
  • Agent Trace Fine-tuning: Benefits from a unique fine-tuning process using high-quality agent traces, enabling it to learn complex vulnerability identification and patching strategies.

Good for

  • Automated Security Patching: Ideal for applications requiring automated identification and fixing of common vulnerabilities and exposures (CVEs).
  • Code Security Analysis: Useful for developers and security researchers looking to integrate advanced vulnerability remediation capabilities into their workflows.
  • Benchmarking Security Agents: Can serve as a strong baseline or component in developing and evaluating new security-focused AI agents.

Finetunes
2 models