README.md

---
language:
- en
- zh
license: apache-2.0
base_model:
- Qwen/Qwen3.5-397B-A17B
pipeline_tag: image-text-to-text
tags:
- agnes-ai
- reasoning
- multimodal
- long-context
- api
---

<p align="center">
  <img width="132" src="assets/agnes_logo.svg" alt="Agnes AI logo">
</p>

<p align="center">
  <a href="https://agnes-ai.com/"><img src="https://img.shields.io/badge/Agnes_AI-Website-3248AF" alt="Agnes AI website"></a>
  <a href="https://wiki.agnes-ai.com/en/docs/agnes-25-pro-alpha"><img src="https://img.shields.io/badge/Agnes_2.5_Pro_Alpha-API_Docs-3248AF" alt="Agnes API docs"></a>
  <a href="https://artificialanalysis.ai/models/agnes-2-5-pro-alpha"><img src="https://img.shields.io/badge/Artificial_Analysis-Model_Page-111827" alt="Artificial Analysis model page"></a>
</p>

# Agnes 2.5 Pro Alpha

Hello! 👋 Today we are introducing **Agnes 2.5 Pro Alpha**, our most capable reasoning model for advanced coding, scientific problem solving, long-context analysis, multimodal understanding, and agentic workflows.

Highlights:

- **Competitive benchmark performance:** Agnes outperforms Qwen3.5-397B on six of the eight evaluations visualized below, with particularly clear gains on Terminal-Bench v2.1, CritPt, and AA-Omniscience Accuracy.
- **Built for demanding work:** a **1M-token context window**, up to **65,536 output tokens**, extended reasoning, tool calling, and text, image understanding.

<img style="width:100%;max-width:1100px" src="assets/agnes_benchmarks.svg" alt="LLM benchmark evaluation comparing Agnes 2.5 Pro Alpha with flagship-scale models" title="Agnes 2.5 Pro Alpha benchmark evaluation">

## Agnes 2.5 Pro Alpha

A multimodal reasoning model available through the Agnes AI API. The model combines long-context understanding with strong coding and scientific reasoning, while retaining the throughput and pricing needed for production workloads.

### Benchmarks

Agnes 2.5 Pro Alpha is evaluated against the same comparison set selected for Ornith-1.0-397B: Qwen3.5-397B, Qwen3.7-Max, GLM-5.2-744B, MiniMax-M3-428B, DeepSeek-V4-Pro-1.6T, Claude Opus 4.7, and Claude Opus 4.8. Every result below is an independent Artificial Analysis benchmark measurement.

<div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;width:100%;margin:0 auto;padding:16px 0;overflow-x:auto">
<table style="width:100%;table-layout:fixed;border-collapse:collapse;font-size:11px;min-width:1100px">
<thead><tr>
<th style="width:20%;padding:10px 6px;text-align:left;border-bottom:2px solid #3248AF;color:#3248AF">Benchmark</th>
<th style="width:10%;padding:10px 6px;text-align:center;font-weight:700;border-bottom:2px solid #3248AF;color:#3248AF;background:rgba(50,72,175,.09)">Agnes 2.5<br>Pro Alpha</th>
<th style="width:10%;padding:10px 6px;text-align:center;border-bottom:2px solid #3248AF">Qwen3.5<br>397B</th>
<th style="width:10%;padding:10px 6px;text-align:center;border-bottom:2px solid #3248AF">Qwen3.7<br>Max</th>
<th style="width:10%;padding:10px 6px;text-align:center;border-bottom:2px solid #3248AF">GLM-5.2<br>744B</th>
<th style="width:10%;padding:10px 6px;text-align:center;border-bottom:2px solid #3248AF">MiniMax-M3<br>428B</th>
<th style="width:10%;padding:10px 6px;text-align:center;border-bottom:2px solid #3248AF">DeepSeek-V4-Pro<br>1.6T</th>
<th style="width:10%;padding:10px 6px;text-align:center;border-bottom:2px solid #3248AF">Claude Opus<br>4.7</th>
<th style="width:10%;padding:10px 6px;text-align:center;border-bottom:2px solid #3248AF">Claude Opus<br>4.8</th>
</tr></thead>
<tbody>
<tr><td colspan="9" style="padding:8px 12px;font-weight:600;color:#3248AF;background:rgba(50,72,175,.07)">Agentic Work &amp; Coding</td></tr>
<tr><td style="padding:7px">GDPval-AA v2 <sup>†</sup></td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">33.8</td><td style="padding:7px;text-align:center">23.2</td><td style="padding:7px;text-align:center">38.6</td><td style="padding:7px;text-align:center">50.3</td><td style="padding:7px;text-align:center">44.3</td><td style="padding:7px;text-align:center">54.5</td><td style="padding:7px;text-align:center">49.5</td><td style="padding:7px;text-align:center">54.2</td></tr>
<tr><td style="padding:7px">τ³-Banking</td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">12.4</td><td style="padding:7px;text-align:center">13.4</td><td style="padding:7px;text-align:center">11.8</td><td style="padding:7px;text-align:center">34.6</td><td style="padding:7px;text-align:center">15.3</td><td style="padding:7px;text-align:center">39.6</td><td style="padding:7px;text-align:center">34.6</td><td style="padding:7px;text-align:center">34.2</td></tr>
<tr><td style="padding:7px">Terminal-Bench v2.1</td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">67.0</td><td style="padding:7px;text-align:center">51.3</td><td style="padding:7px;text-align:center">74.5</td><td style="padding:7px;text-align:center">77.9</td><td style="padding:7px;text-align:center">65.2</td><td style="padding:7px;text-align:center">78.7</td><td style="padding:7px;text-align:center">83.1</td><td style="padding:7px;text-align:center">84.6</td></tr>
<tr><td style="padding:7px">SciCode</td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">42.2</td><td style="padding:7px;text-align:center">42.0</td><td style="padding:7px;text-align:center">48.8</td><td style="padding:7px;text-align:center">50.5</td><td style="padding:7px;text-align:center">45.4</td><td style="padding:7px;text-align:center">49.2</td><td style="padding:7px;text-align:center">54.5</td><td style="padding:7px;text-align:center">53.5</td></tr>
<tr><td colspan="9" style="padding:8px 12px;font-weight:600;color:#3248AF;background:rgba(50,72,175,.07)">Long Context &amp; Scientific Reasoning</td></tr>
<tr><td style="padding:7px">AA-LCR</td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">73.0</td><td style="padding:7px;text-align:center">72.7</td><td style="padding:7px;text-align:center">74.7</td><td style="padding:7px;text-align:center">76.7</td><td style="padding:7px;text-align:center">80.3</td><td style="padding:7px;text-align:center">75.3</td><td style="padding:7px;text-align:center">75.3</td><td style="padding:7px;text-align:center">73.0</td></tr>
<tr><td style="padding:7px">Humanity's Last Exam</td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">33.6</td><td style="padding:7px;text-align:center">29.0</td><td style="padding:7px;text-align:center">40.5</td><td style="padding:7px;text-align:center">41.1</td><td style="padding:7px;text-align:center">39.0</td><td style="padding:7px;text-align:center">41.0</td><td style="padding:7px;text-align:center">42.3</td><td style="padding:7px;text-align:center">48.7</td></tr>
<tr><td style="padding:7px">GPQA Diamond</td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">87.6</td><td style="padding:7px;text-align:center">89.3</td><td style="padding:7px;text-align:center">92.3</td><td style="padding:7px;text-align:center">89.5</td><td style="padding:7px;text-align:center">92.9</td><td style="padding:7px;text-align:center">92.8</td><td style="padding:7px;text-align:center">91.4</td><td style="padding:7px;text-align:center">92.0</td></tr>
<tr><td style="padding:7px">CritPt</td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">10.9</td><td style="padding:7px;text-align:center">1.7</td><td style="padding:7px;text-align:center">13.4</td><td style="padding:7px;text-align:center">20.9</td><td style="padding:7px;text-align:center">3.7</td><td style="padding:7px;text-align:center">18.0</td><td style="padding:7px;text-align:center">12.0</td><td style="padding:7px;text-align:center">20.9</td></tr>
<tr><td colspan="9" style="padding:8px 12px;font-weight:600;color:#3248AF;background:rgba(50,72,175,.07)">Knowledge Reliability</td></tr>
<tr><td style="padding:7px">AA-Omniscience Accuracy</td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">33.5</td><td style="padding:7px;text-align:center">30.8</td><td style="padding:7px;text-align:center">31.1</td><td style="padding:7px;text-align:center">24.3</td><td style="padding:7px;text-align:center">16.7</td><td style="padding:7px;text-align:center">49.1</td><td style="padding:7px;text-align:center">48.9</td><td style="padding:7px;text-align:center">48.8</td></tr>
<tr><td style="padding:7px">Non-Hallucination Rate</td><td style="padding:7px;text-align:center;font-weight:700;color:#3248AF;background:rgba(50,72,175,.05)">11.9</td><td style="padding:7px;text-align:center">11.1</td><td style="padding:7px;text-align:center">74.4</td><td style="padding:7px;text-align:center">73.7</td><td style="padding:7px;text-align:center">81.6</td><td style="padding:7px;text-align:center">5.2</td><td style="padding:7px;text-align:center">57.7</td><td style="padding:7px;text-align:center">60.7</td></tr>
</tbody>
</table>
</div>

<p style="font-size:11px;opacity:.72">
† GDPval-AA v2 uses Artificial Analysis' normalized score, <code>(Elo − 500) / 2000</code>. Non-Hallucination Rate is <code>1 − hallucination rate</code>. Higher is better for every benchmark. Snapshot checked August 18, 2026; values may change as evaluations are updated.
</p>

## Model Information

| Property | Value |
|---|---|
| Developed by | Agnes AI |
| Model name | Agnes 2.5 Pro Alpha |
| Model type | Multimodal reasoning model |
| License | Apache License 2.0 |
| Languages | English, Chinese |
| Context window | 1,048,576 tokens |
| Maximum output | 65,536 tokens |
| Input modalities | Text, image |
| Output modality | Text |
| Precision | BF16 |
| Tool calling | Yes |
| Streaming | Yes |
| Release date | July 2026 |

## License

This repository is licensed under the [Apache License 2.0](LICENSE).

Agnes 2.5 Pro Alpha is a post-trained derivative of [Qwen/Qwen3.5-397B-A17B](https://huggingface.co/Qwen/Qwen3.5-397B-A17B), which is also licensed under Apache License 2.0. Original copyright notices are retained. Additional post-training was performed by Agnes AI.

See the `LICENSE` file in this repository for the full terms.

## Hardware Requirements

The Quickstart launch command uses 8-GPU tensor parallelism. The checkpoint is a large multi-shard BF16 package; a single GPU is not sufficient.

| Resource | Recommendation |
|---|---|
| GPUs | 8× NVIDIA H200 (141 GB) or equivalent |
| Tensor parallel | `--tp 8` |
| Host memory / disk | Fast NVMe with about **1 TB** free for weights, tokenizer files, and download cache |
| Context length | The sample command sets `--context-length 1024000`. If you hit out-of-memory errors, lower this value |
| Network | Optional. The same model is also served at `https://apihub.agnes-ai.com/v1` without local GPUs |

## Quickstart

<div style="border-left:4px solid #3248AF;background:rgba(50,72,175,.08);border-radius:6px;padding:12px 16px;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;font-size:14px;line-height:1.6">
<div style="font-weight:700;color:#3248AF;margin-bottom:6px">REASONING MODEL</div>
<p style="margin:0"><b>Agnes 2.5 Pro Alpha</b> uses extended reasoning for complex tasks and is available through OpenAI-compatible Chat Completions and Responses APIs. Keep API keys in environment variables and use publicly accessible URLs for image inputs.</p>
</div>

<h3>SGLang</h3>

```bash
python -m sglang.launch_server \
    --model-path Agnes-AI/Agnes-2.5-Pro-Alpha \
    --served-model-name agnes-2.5-pro-alpha \
    --tp 8 \
    --host 0.0.0.0 --port 8000 \
    --context-length 1024000 \
    --mem-fraction-static 0.85 \
    --tool-call-parser qwen3_coder \
    --reasoning-parser qwen3
```

### Chat Completions

```bash
export AGNES_API_KEY="your-api-key"

curl https://apihub.agnes-ai.com/v1/chat/completions \
  -H "Authorization: Bearer ${AGNES_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "agnes-2.5-pro-alpha",
    "messages": [
      {
        "role": "user",
        "content": "Review this API handler for security issues and provide a corrected version."
      }
    ],
    "temperature": 1.0,
    "max_tokens": 2000
  }'
```

### Python

```bash
pip install openai
```

```python
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["AGNES_API_KEY"],
    base_url="https://apihub.agnes-ai.com/v1",
)

response = client.chat.completions.create(
    model="agnes-2.5-pro-alpha",
    messages=[
        {
            "role": "user",
            "content": "Design a fault-tolerant event processing architecture.",
        }
    ],
    temperature=1.0,
    max_tokens=2000,
)

print(response.choices[0].message.content)
```

### Image Understanding

```python
response = client.chat.completions.create(
    model="agnes-2.5-pro-alpha",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Explain this chart and call out anomalies."},
                {
                    "type": "image_url",
                    "image_url": {"url": "https://example.com/chart.png"},
                },
            ],
        }
    ],
)
```

### Responses API

```bash
curl https://apihub.agnes-ai.com/v1/responses \
  -H "Authorization: Bearer ${AGNES_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "agnes-2.5-pro-alpha",
    "input": "Create a step-by-step migration plan from a monolith to services.",
    "max_output_tokens": 2000
  }'
```

## Recommended Inference Settings

Use sampling rather than greedy decoding. Leave enough `max_tokens` / `max_output_tokens` for extended reasoning.

| Setting | Recommended |
|---|---|
| `temperature` | 1.0 |
| `top_p` | 0.95 |
| `top_k` | 20 |
| `repetition_penalty` | 1.05 |
| `max_tokens` | 2000 or higher |

Raise `max_tokens` if a response stops early.

## Model Capabilities

| Capability | Support |
|---|---|
| Advanced reasoning | Yes |
| Coding and debugging | Yes |
| Long-context analysis | 1M tokens |
| Maximum output | 65,536 tokens |
| Image understanding | Yes, via public image URL |
| Tool calling | Yes |
| Streaming | Yes |
| OpenAI-compatible APIs | Chat Completions and Responses |

Agnes 2.5 Pro Alpha is especially well suited to repository-level coding, technical research, document synthesis, visual analysis, and tool-enabled agents that need to reason across long and complex contexts.

## Responsible Use

Model outputs can contain errors. Validate high-impact decisions and tool actions in the application layer, and review the applicable Agnes AI service terms before sending sensitive or regulated data.

## Citation

```bibtex
@misc{agnes25proalpha2026,
  title        = {Agnes 2.5 Pro Alpha},
  author       = {{Agnes AI}},
  year         = {2026},
  month        = jul,
  howpublished = {API model},
  url          = {https://wiki.agnes-ai.com/en/docs/agnes-25-pro-alpha}
}
```