tsinghua-sigs-robot-lab/VeriLoop-Coder-E1

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

VeriLoop Coder-E1 is a 27 billion parameter model backend developed by Tsinghua SIGS Robot Lab, built on the Qwen3.6-27B backbone with a 32768 token context length. It is the open model component of the VeriLoop coding-intelligence system architecture, designed for code reasoning, repository understanding, and artifact generation. This model is distinguished by its integration with a non-public 'Self-Harness' control plane and 'Surface Host PEFT' adapters, which enable evidence-governed correction and specialized behavioral control for robust software engineering tasks, making it particularly strong for tool-mediated code generation and repository repair.

Loading preview...

VeriLoop Coder-E1: An Open Model for Evidence-Governed Software Engineering

VeriLoop Coder-E1, developed by Tsinghua SIGS Robot Lab, is a 27 billion parameter model built on the Qwen3.6-27B backbone, designed as the open model backend for the broader VeriLoop coding-intelligence system. It specializes in code reasoning, repository understanding, and artifact generation, operating within a 32768 token context window. This model is unique due to its architectural design, which integrates with a non-public "Self-Harness" control plane and utilizes "Surface Host PEFT" adapters for enhanced behavioral control.

Key Capabilities

  • Code Reasoning & Generation: Excels at understanding codebases, generating code, and drafting patches.
  • Repository Understanding: Capable of navigating and interpreting complex software repositories.
  • Evidence-Governed Correction: Designed to work within a system that uses evidence to drive correction, validation, and repair, rather than simple retries.
  • Specialized Behavioral Control: Leverages four narrow-domain PEFT adapters (ToolSpec, Uncertainty, Rollback, Evidence Binding) to strengthen tool-call discipline, contract adherence, and evidence-sensitive reasoning.
  • Agent-System Integration: While the full agent runtime (Voder) and production Self-Harness are private, Coder-E1 provides the model-side foundation for advanced coding agents.
  • Robust Error Handling: Emphasizes deterministic-generative separation and failure-aware convergence, allowing the system to distinguish various failure types and perform surgical repairs.

Good For

  • Tool-Mediated Software Engineering Agents: Ideal as the core model for building sophisticated coding agents that require robust interaction with tools and environments.
  • Repository Repair & Bug Localization: Strong capabilities for identifying and surgically repairing bugs and managing cross-file API changes.
  • Validator-Aware Workflows: Suited for scenarios where code generation needs to be tightly coupled with validation and verification processes.
  • Benchmark & Evaluation Research: Provides a strong foundation for research into coding agent performance and evidence-based evaluation.
  • Long-Running Engineering Tasks: Supports tasks requiring checkpoints and auditability through its evidence-governed approach.

Finetunes
1 models