OpenCSG Agentic-27B: 27.36B-Parameter Tool-Use Model Runs on One A800, Commercial Use Needs Permission

OpenCSG has released Agentic-27B, a downloadable model with about 27.36 billion parameters and BF16 weights, built for multi-step tool use and private enterprise deployments. Commercial use requires authorization under the OpenCSG Model and Dataset License Agreement.

OpenCSG's technical introduction and model card are written in Chinese. They describe a focus on selecting Skills, constructing tool arguments, recovering from failures and checking whether an assigned task is actually complete.

Base model and training

Agentic-27B is the second public version in OpenCSG's Agentic series. It replaces the earlier Qwen3-30B-A3B mixture-of-experts base with the dense Qwen3.8-27B model.

OpenCSG says it trained with low-rank adaptation and direct preference optimization (LoRA DPO) on execution records generated through its platform. The preference dataset grew from roughly 3,000 examples to 5,857. According to OpenCSG, the samples include tool-use trajectories, low-scoring answers, failed calls and verification feedback, and the records were redacted, checked and screened before training.

Benchmark results (company-reported)

All reported benchmark results come from OpenCSG. They are not independent certification of reliability in a company's own systems.

• Agentic Eval average: about 0.83, second among the models in OpenCSG's run.
• Suite size: 119 tasks, each attempted three times. The detailed model card says only 103 tasks received valid scores.
• Passed all three attempts: 74 of the 103 scoreable tasks.
• Passed at least once: 96 of the 103 scoreable tasks.
• Previous model: OpenCSG compares the all-three count with 11 for its earlier model.

These are task counts, not percentages. Because scoring coverage differs, comparisons between releases should be read with that in mind.

Deployment figures

OpenCSG reports running the model on a single NVIDIA A800 with 80GB of memory. In its test, BF16 weights loaded using 51.1 GiB, with a 32,768-token serving context and up to eight concurrent requests. The model configuration supports a maximum of 262,144 positions, so the tested context is smaller than the configured maximum.

License and safety guidance

The Chinese-language model card qualifies 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.

Specs at a glance

Item	Detail
Parameters	About 27.36B (dense)
Weights	BF16
Base model	Qwen3.8-27B
Fine-tuning	LoRA DPO, 5,857 preference examples
Configured max positions	262,144
Tested serving context	32,768 tokens
Pricing	Not yet disclosed

What this means

For teams evaluating private agent deployments, the main practical detail is hardware: OpenCSG reports that the BF16 model fits on one 80GB A800 with room for eight concurrent requests at a 32,768-token context. Longer contexts than that were not part of the reported test.

The license matters as much as the weights. Anyone planning to build a commercial agent, derived model or API product on Agentic-27B needs authorization from OpenCSG first.

The benchmark evidence is limited. The scores are self-reported on OpenCSG's own suite, and 16 of the 119 tasks did not receive valid scores. The jump from 11 to 74 all-three-attempt passes is OpenCSG's comparison with its previous model, and the scoring coverage differs, so it should not be treated as a like-for-like measure. Teams should test the model on their own tool workflows, with the human confirmation and least-privilege controls OpenCSG recommends for write operations.