File size: 3,044 Bytes
bbe5330
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cf1c403
bbe5330
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
84
85
86
---
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:
- document-understanding
- knowledge-work
- pdf
---

# AIBuildAI ProWork-PDF-27B

## Introduction

**AIBuildAI ProWork-PDF-27B** is Qwen3.8-27B post-trained by AIBuildAI on our own data for professional work on long documents: reading a working professional's request and the PDF packet it refers to (records, tables, charts, plans and regulations, often more than fifty pages), working out the figures from the packet, and answering in writing, as a memo, a note or a table.

- **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

GDP.pdf (Surge AI's 100 public tasks of document work across ten domains), run under Artificial Analysis's protocol: the document's text and page images in the prompt, every rubric criterion judged by GPT-5.6 Luna (medium). All-pass is the share of attempts that meet every criterion.

![All-pass on GDP.pdf](assets/gdppdf_allpass.png)

<sub>Lighter bars are Artificial Analysis's published scores; the other bars are our own runs, the base, Qwen3.8-27B, through its hosted API.</sub>

AIBuildAI ProWork-PDF-27B reaches 21.0% all-pass, 5.4 points over its base, and comes within 0.6 points of GPT-6 Luna (max).

## Quickstart

Serve with vLLM:

```bash
vllm serve aibuildai-inc/ProWork-PDF-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, giving the document's text and its page images:

```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-PDF-27B",
    messages=[{"role": "user", "content": [
        {"type": "text", "text": "Using the attached annual report, list each business segment with its revenue and operating margin, in a table. ..."},
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
    ]}],
    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. Give the model the document's text together with its page images, so it can read charts, plans and scanned tables.

## 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{aibuildai2026proworkpdf,
  title  = {AIBuildAI ProWork-PDF-27B},
  author = {{AIBuildAI}},
  year   = {2026},
  url    = {https://huggingface.co/aibuildai-inc/ProWork-PDF-27B}
}
```