--- license: apache-2.0 base_model: Qwen/Qwen3.8-27B base_model_relation: finetune library_name: transformers pipeline_tag: image-text-to-text language: - en tags: - agent - knowledge-work - tool-use --- # AIBuildAI ProWork-27B ## Introduction **AIBuildAI ProWork-27B** is Qwen3.8-27B post-trained by AIBuildAI on our own data for agentic professional work: reading a working professional's request and its attached records, data and regulations, doing the analysis with tools (web search, file reading, code), and handing back a finished deliverable such as a memo, a spreadsheet model or a slide deck. - **Base:** [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B), 27B parameters, text and image input, 262,144-token context. - **Weights:** bf16. ## Evaluation GDPval (OpenAI's 220 public tasks of economically valuable knowledge work), 3 attempts per task. Each model's deliverables are compared head to head with those of Qwen3.8-27B (xhigh), fixed at its Artificial Analysis leaderboard Elo of 1427, by a panel of three LLM judges (GPT, Gemini and Claude). Whiskers are 95% task-bootstrap intervals. ![Elo score on GDPval](assets/gdpval_elo.png) Our own implementation of Artificial Analysis-style GDPval Elo; not leaderboard scores. AIBuildAI ProWork-27B gains 125 Elo over its base and comes within 25 Elo of GPT-6 Astra (max). ## Quickstart Serve with vLLM: ```bash vllm serve aibuildai-inc/ProWork-27B \ --max-model-len 262144 \ --reasoning-parser qwen3 \ --enable-auto-tool-choice --tool-call-parser qwen3_xml \ --speculative-config '{"method": "mtp", "num_speculative_tokens": 3}' ``` Then call it with any OpenAI-compatible client: ```python from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY") response = client.chat.completions.create( model="aibuildai-inc/ProWork-27B", messages=[{"role": "user", "content": "Draft a one-page memo comparing the two vendor quotes below ..."}], temperature=1.0, top_p=0.95, max_tokens=65536, extra_body={"top_k": 20}, ) print(response.choices[0].message.content) ``` **Recommended settings:** thinking on, temperature 1.0, top-p 0.95, top-k 20, up to 65,536 output tokens per call. For agentic use, give the model tools for web search, reading files and running code, and allow long sessions; the evaluation above allowed up to 250 turns. ## Limitations As with any model, check figures, citations and legal or medical statements before relying on them. ## License Apache 2.0, the license of the base model Qwen3.8-27B. ## Citation ```bibtex @misc{aibuildai2026prowork, title = {AIBuildAI ProWork-27B}, author = {{AIBuildAI}}, year = {2026}, url = {https://huggingface.co/aibuildai-inc/ProWork-27B} } ```