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, developed by Tsinghua SIGS Robot Lab, is a 27 billion parameter model backend built on Qwen3.6-27B, designed for code reasoning, repository understanding, and artifact generation. It features detachable PEFT-based Surface Host Adapters for specialized behavioral control in areas like tool specification, uncertainty, rollback, and evidence binding. This model is a core component of the VeriLoop coding intelligence system, optimized for evidence-governed correction and robust software engineering tasks.

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 based on Qwen3.6-27B, serving as the open model backend for the VeriLoop coding intelligence system. It is specifically designed for advanced code reasoning, repository understanding, and artifact generation, distinguishing itself through an architecture that emphasizes "evidence-governed correction" rather than simple prompt accumulation.

Key Capabilities & Differentiators

  • Modular Adaptation: Utilizes detachable PEFT-based Surface Host Adapters for specialized behavioral control, including ToolSpec, Uncertainty, Rollback, and Evidence Binding, enhancing discipline in tool calls, contract adherence, and validation-aware repair.
  • System-Level Performance: Achieves strong benchmark results (e.g., 85.20% on SWE-bench Verified, 62.38% on SWE-bench Pro) when integrated within its prototype agent runtime and pre-release Self-Harness control loop, highlighting its effectiveness in a complete system context.
  • Robust Engineering Protocol: Incorporates a 14-rule public Self-Harness contract, emphasizing current-task supremacy, evidence-based escalation, surgical repair, and deterministic enforcement for auditable software engineering processes.
  • Failure-Aware Convergence: Distinguishes various failure types (invalid candidate, missing dependency, etc.) to enable precise, targeted repairs and safe stopping when evidence does not support delivery.

Good For

  • Software Engineering Agents: Ideal for building tool-mediated coding agents, especially those requiring robust validation and repair workflows.
  • Codebase Navigation & Repair: Excels in repository understanding, bug localization, surgical patch drafting, and cross-file API changes.
  • Auditable Workflows: Supports long-running engineering tasks with checkpointing and traceability for verifiable evidence chains.
  • Benchmark Research: Useful for research into coding agent benchmarks and evaluation methodologies, particularly those focusing on system-level performance.

Finetunes
1 models