ProCreations/grug-27b-v2

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

ProCreations/grug-27b-v2 is a 27 billion parameter language model, fine-tuned from Qwen 3.8 27B, specifically optimized for compact reasoning, code generation, and tool use. It features a native Multi-Token Prediction (MTP) draft head integrated directly into the checkpoint for faster inference. This model demonstrates improved performance on coding benchmarks like HumanEval and MBPP, as well as SWE-bench tasks, while offering configurable reasoning effort levels. It is designed for developers needing an efficient and capable model for programming and agentic workflows with a 32768 token context length.

Loading preview...

ProCreations/grug-27b-v2: Compact Code and Tooling

ProCreations/grug-27b-v2 is a 27 billion parameter model built upon the Qwen 3.8 27B foundation, specifically fine-tuned for efficient reasoning, robust code generation, and effective tool utilization. A key differentiator is its integrated native Multi-Token Prediction (MTP) draft head, which is part of the main checkpoint, eliminating the need for separate draft models and enhancing inference speed.

Key Capabilities and Performance

  • Enhanced Code Generation: grug-27b-v2 shows significant improvements over its predecessor (v1.1) and strong performance against the base Qwen 3.8 27B on coding benchmarks. It achieves 97.0% on Medium HumanEval, 90.7% on Medium MBPP, and 88.0% on Medium BFCL subset.
  • Advanced Tool Use: The model excels in agentic workflows, fixing 10/12 issues on a SWE-bench Verified Django/SymPy subset with medium effort, and 11/12 with extended context. It also resolves a critical session-title bug, passing 90/90 checks for tool-call leaks.
  • Configurable Reasoning Effort: Users can adjust the model's reasoning verbosity via chat_template_kwargs.reasoning_effort (low, medium, xhigh), allowing for a trade-off between thought length and task performance. It uses significantly fewer reasoning tokens than the base Qwen model.
  • Faster Inference with MTP: The integrated MTP head boosts generation speed, achieving up to 142.44 tokens/sec with two draft tokens in specific workloads, compared to 68.02 tokens/sec without MTP.

Use Cases and Considerations

This model is well-suited for applications requiring efficient code generation, automated bug fixing, and complex tool-use scenarios. While it demonstrates strong coding and tool-use capabilities, it trades some raw accuracy in areas like mathematics (e.g., 84.7% on a medium MATH subset compared to Qwen's 95.3%) for its compact reasoning and specialized focus. The model maintains the Apache-2.0 license of its foundation.

Finetunes
2 models