OpenCSG Releases Agentic-27B for Tool Workflows, With Commercial Use Requiring Permission

OpenCSG has released Agentic-27B, a downloadable model designed for multi-step tool use and private enterprise deployments. Its Chinese technical introduction describes a focus on selecting Skills, constructing tool arguments, recovering from failures and checking whether an assigned task is actually complete.

The release is the second public version in the company's Agentic series. It replaces the earlier Qwen3-30B-A3B mixture-of-experts base with the dense Qwen3.8-27B model. OpenCSG says it used low-rank adaptation and direct preference optimization, or LoRA DPO, on execution records generated through its platform.

The preference-training dataset grew from roughly 3,000 examples to 5,857. OpenCSG says those samples include tool-use trajectories, low-scoring answers, failed calls and verification feedback, with records redacted, checked and screened before training. The model has about 27.36 billion parameters and BF16 weights.

All reported benchmark results come from OpenCSG. In its Agentic Eval, the company reports an average score of about 0.83, placing second among the models in that run. The suite contains 119 tasks, each attempted three times, but the detailed model card says only 103 tasks received valid scores.

On those 103 scoreable tasks, 74 passed all three attempts, while 96 passed at least once. OpenCSG compares the all-three success count with 11 for its previous model. These are task counts, not percentages; the differing scoring coverage should be considered when comparing releases. The results are not independent certification of reliability in a company's own systems.

OpenCSG also reports deployment on a single NVIDIA A800 with 80GB of memory. Its test loaded BF16 weights using 51.1 GiB, with a 32,768-token serving context and up to eight concurrent requests. That test context is smaller than the model configuration's 262,144-position maximum.

The Chinese-language model card makes a significant qualification to the open release: commercial use requires authorization under the OpenCSG Model and Dataset License Agreement. Downloadable weights do not provide unrestricted commercial permission, including for derived models, agents or API products. OpenCSG also says sensitive write operations need human confirmation and least-privilege controls.

SEE ALSO: DeepSeek Harness Plans an Official Plugin Marketplace