File size: 2,784 Bytes
d872155
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
---
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)

<sub>Our own implementation of Artificial Analysis-style GDPval Elo; not leaderboard scores.</sub>

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}
}
```