OpenCSG presents the second public version of its Agentic series in a model card on Hugging Face. Pandaily reported on the model on 7 November, drawing on the model card and the company’s Chinese-language technical overview.

Agentic-27B is designed for AI agents, programs that carry out work tasks using a model. The model card says the model is tuned to choose the right skill, a packaged set of instructions and tools for a task, to form valid tool calls, to read the results that tools return and to recover from errors.

The model is based on Qwen3.8-27B, which has about 27.36 billion parameters. Its predecessor was based on Qwen3-30B-A3B, a mixture-of-experts (MoE) model that uses only some of its parameters at each computation step.

OpenCSG fine-tuned the model with LoRA DPO. In DPO, a model learns from pairs of answers to prefer the better one; LoRA trains only a small add-on rather than the whole model. The training data comprised 5,857 samples, up from about 3,000 in the previous version. OpenCSG collected the samples from use of its platform, for example low-scoring answers and failed tool calls.

The results are the company’s own

OpenCSG evaluated the model on its own Agentic Eval benchmark, running each of its 119 agent tasks three times. Task completion accounts for 80 per cent of the score and reliability for 20 per cent, and the total is multiplied by a safety score. A run passes with a score of at least 0.75. OpenCSG also used a judge model in scoring.

Only 103 of the 119 tasks produced scores, and their mean score was 0.83. All three runs succeeded on 74 tasks and at least one run on 96. The model card’s highlights present the latter figure as 96/119 and compare it with the predecessor’s 37/119. The predecessor’s mean is given as 0.46.

Pandaily reported that OpenCSG compares the number of tasks passed in all three runs with 11 for the predecessor, and says the model ranked second among the models tested. Pandaily noted that any comparison between versions must take into account the different number of scored tasks. Nor, it added, do the results prove that the model would work reliably in a company’s own systems.

Arena-Hard-v2.0 measures how well a model follows difficult prompts and writes creatively. The model’s hard_prompt score rose from the predecessor’s 26.6 to 57.68 and its creative_writing score from 48.2 to 68.39.

One GPU was enough in testing

OpenCSG ran the model on the vLLM inference server on a single NVIDIA A800 GPU with 80 gigabytes of memory. The BF16 weights took up 51.1 GiB. Memory use depends on factors including context length and the vLLM version, the company said. The test used a context window of 32,768 tokens, although the model’s configuration sets a maximum length of 262,144 positions.

A single request generated about 29 tokens per second; eight concurrent requests generated about 213 in total.

The model card says the model’s outputs should not be used unchecked for medical, legal or financial decisions or for actions that require broad permissions. Sensitive changes require human confirmation and the narrowest possible access rights.

The weights can be downloaded from Hugging Face under OpenCSG’s model and dataset licence agreement. Commercial use requires the company’s permission, requested by email, including when a model, system, agent, API service or product has been trained on or enhanced with Agentic-27B. The model card includes a launch example for vLLM.