opencsg/Agentic-27B

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 11, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

OpenCSG Agentic-27B is a 27.36 billion parameter language model developed by OpenCSG, based on Qwen/Qwen3.8-27B. It is specifically optimized for agentic capabilities, tool use, and private execution scenarios, featuring a 32,768 token context length. The model excels at multi-step task completion, understanding system constraints, and handling complex workflows, making it suitable for enterprise-grade agent deployments.

Loading preview...

OpenCSG Agentic-27B: An Agent Model for Real-World Tasks

OpenCSG Agentic-27B is the second public release in the OpenCSG Agentic series, built upon the Qwen/Qwen3.8-27B dense base model. It has been fine-tuned using LoRA DPO on a significantly expanded dataset (5,857 samples, nearly double the previous version) generated by the OpenCSG platform's data flywheel. This model is engineered not for single-turn Q&A, but for robust performance in real working environments, focusing on complex, multi-step task execution.

Key Capabilities and Enhancements

  • Superior Agent Execution: Achieves an Agentic Eval mean score of 0.83 (up from 0.46) and pass@3 of 96/119 (up from 37/119), demonstrating significant improvements in handling multi-step tool-use tasks.
  • Enhanced General Performance: Simultaneously boosts performance on general hard prompts and creative writing, with hard_prompt scores improving from 26.6 to 57.68 and creative_writing from 48.2 to 68.39 on Arena-Hard-v2.0.
  • Real Tool Execution Data Loop: Training data emphasizes tool selection, argument construction, result reading, failure recovery, and final verification, ensuring practical applicability.
  • Single-GPU Deployability: The BF16 weights can be loaded on a single NVIDIA A800 80GB GPU, requiring 51.1 GiB of memory, making it accessible for private deployments.
  • Long Context Window: Supports a vLLM context of 32,768 tokens, enabling processing of extensive inputs and complex task sequences.

Ideal Use Cases

  • Enterprise Agent Deployments: Suited for local agents, platform Skills, structured tool use, and cross-system workflows within private enterprise environments.
  • Complex Task Automation: Excels in scenarios requiring understanding system constraints, selecting appropriate skills, generating valid parameters, and handling execution feedback.
  • Tool-Augmented Applications: Designed for applications that leverage external tools and APIs, focusing on robust interaction and verifiable outcomes.

Finetunes
1 models