OpenCSG has released Agentic-27B, a downloadable model intended for private enterprise deployments and multi-step tool use. The company describes the model's focus as selecting Skills, constructing tool arguments, recovering from failures, and verifying if a task is complete. This release marks the second public version in the company's Agentic series.
The new model replaces the previous Qwen3-30B-A3B mixture-of-experts base with a dense Qwen3.8-27B model. OpenCSG states it utilized low-rank adaptation and direct preference optimization, or LoRA DPO, on execution records from its own platform. The dataset for preference training increased from approximately 3,000 examples to 5,857. According to OpenCSG, these samples include tool-use trajectories, failed calls, verification feedback, and low-scoring answers, all of which were redacted, checked, and screened before training.
The model features about 27.36 billion parameters with BF16 weights. OpenCSG reported an average score of about 0.83 in its Agentic Eval, which it claims placed the model second among the models in that specific run. The evaluation suite consisted of 119 tasks attempted three times each, though the detailed model card notes that only 103 tasks received valid scores. Of those 103 scoreable tasks, 74 passed all three attempts and 96 passed at least once. OpenCSG notes that the previous model had an all-three success count of 11. The company specifies these are task counts rather than percentages and notes that differing scoring coverage should be considered when comparing the two releases.
OpenCSG states that these results are not independent certification of reliability in a company's own systems. In a deployment test on a single NVIDIA A800 with 80GB of memory, the company reported that BF16 weights loaded using 51.1 GiB. This test featured a 32,768-token serving context and up to eight concurrent requests, which is smaller than the model's maximum configuration of 262,144 positions.
The Chinese-language model card includes a significant qualification regarding the open release: commercial use requires authorization under the OpenCSG Model and Dataset License Agreement. OpenCSG states that downloadable weights do not provide unrestricted commercial permission, including for derived models, agents, or API products. Additionally, the company states that sensitive write operations require human confirmation and least-privilege controls.
