tsinghua-sigs-robot-lab/veriloop-coder-e1

VISIONConcurrent 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, developed by Tsinghua SIGS Robot Lab, is a 27 billion parameter model backend for code reasoning, repository understanding, and artifact generation, built on a Qwen3.6-27B backbone. It is designed as an open component of the VeriLoop coding-intelligence system, specializing in software engineering tasks through evidence-governed correction and a unique Surface Host Adapter with narrow-domain PEFT surfaces. This model excels at tasks like bug localization, patch drafting, and tool-mediated code generation, integrating with a non-public Self-Harness for advanced agentic behavior. It achieves strong performance on coding benchmarks such as SWE-bench Verified (85.20%) and Terminal-Bench 2.0 (76.40%) when evaluated as part of a complete agent system.

Loading preview...

VeriLoop Coder-E1: An Evidence-Governed Coding Intelligence Backend

VeriLoop Coder-E1, developed by Tsinghua SIGS Robot Lab, is a 27 billion parameter model built on a Qwen3.6-27B backbone, serving as the open model backend for the VeriLoop coding-intelligence system. It is specifically designed for code reasoning, repository understanding, and artifact generation. A key differentiator is its architecture, which includes a unique Surface Host Adapter with four narrow-domain PEFT surfaces (ToolSpec, Uncertainty, Rollback, Evidence Binding) that enhance sensitivity to tool contracts, uncertainty, state restoration, and evidence alignment. These surfaces are designed to interoperate with a non-public Self-Harness control plane, which governs task compilation, evidence routing, tool governance, and recursive self-evolution.

Key Capabilities

  • Code Reasoning & Generation: Excels at understanding codebases and generating artifacts like patches, source files, and configurations.
  • Evidence-Governed Correction: Utilizes a 14-rule public Self-Harness contract for disciplined engineering, focusing on falsification, surgical repair, and re-verification.
  • Agentic Integration: Designed to be the model component of a larger coding agent system (Voder, currently in development), enabling advanced capabilities like tool-mediated code generation and repository repair.
  • Benchmark Performance: Achieves notable results in system-level evaluations, including 85.20% on SWE-bench Verified and 76.40% on Terminal-Bench 2.0, demonstrating its effectiveness within an agent runtime.

Good For

  • Software Engineering Agents: Building sophisticated coding agents that require robust code understanding and generation.
  • Bug Localization & Repair: Identifying and surgically fixing bugs within codebases.
  • Repository Understanding: Navigating and comprehending complex code repositories.
  • Auditable Workflows: Tasks requiring checkpoints, traceability, and evidence-grounded coding assistance.

Finetunes
1 models