agnes-3.0-qwen

Hello! 👋 Today we are introducing agnes-3.0-qwen, our most capable reasoning model for agentic coding, tool use, scientific problem solving, long-context analysis, and multimodal understanding.

Highlights:

• Strong agentic performance: agnes-3.0-qwen scores highest in its comparison set on τ³-Banking, and is within about one point of the best model on Terminal-Bench 2.1, while being a much smaller model than most of the flagship systems it is compared against.
• Reliable instruction following: on IFBench, agnes-3.0-qwen outperforms GPT-5.6 Sol (Max) and Claude Opus 4.8 (Max).
• Built for demanding work: a 1M-token context window, extended reasoning with selectable reasoning effort (xhigh, medium, low), tool calling, and text, image, and video understanding.

agnes-3.0-qwen

A multimodal reasoning model available as open weights. The model is a hybrid linear-attention / full-attention Mixture-of-Experts architecture (512 experts, 10 active per token) with a multi-token-prediction head, combining long-context understanding with strong coding, agentic, and scientific reasoning at the throughput and cost needed for production workloads.

Benchmarks

agnes-3.0-qwen is compared against GPT-5.6 Sol (Max), Gemini 3.8 Flash, DeepSeek V4 Flash (Max), MiniMax-M3, Claude Opus 4.8 (Max), and Qwen3.8 Max (0902). Results for agnes-3.0-qwen are self-reported on publicly available benchmarks; comparison figures are taken from public reports.

Benchmark	agnes-3.0-qwen	GPT-5.6 Sol (Max)	Gemini 3.8 Flash	DeepSeek V4 Flash (Max)	MiniMax M3	Claude Opus 4.8 (Max)	Qwen3.8 Max (0902)
Agentic Work & Coding
GDP.pdf	16.0	27.2	21.0	12.8	9.8	22.8	22.8
τ³-Banking	48.0	44.3	44.9	39.4	15.3	34.2	47.8
Terminal-Bench 2.1	87.8	88.0	87.6	78.7	65.2	84.6	88.8
Terminal-Bench 4.0	26.1	39.9	19.7	12.1	2.0	21.7	38.9
SciCode	56.1	57.1	56.6	51.9	47.1	54.4	52.1
Knowledge, Reasoning & Instruction Following
GPQA Diamond	91.1	94.1	95.3	90.8	92.9	92.8	92.8
MMLU-Pro	86.2	89.1	90.7	88.2	84.6	89.6	88.8
IFBench	79.6	72.7	—	79.2	82.9	62.2	82.8

Higher is better for every benchmark. agnes-3.0-qwen figures are self-reported; comparison model figures are taken from publicly available reports and may change as evaluations are updated. "—" indicates no published result. Snapshot checked October 10, 2026.

Model Information

Property	Value
Developed by	Agnes AI
Model name	agnes-3.0-qwen
Model type	Multimodal reasoning model (hybrid attention MoE)
Total parameters	~180B (512 experts, 10 active per token)
License	Qwen Community License 1.0
Languages	English, Chinese
Context window	1,048,576 tokens
Maximum output	65,536 tokens
Input modalities	Text, image, video
Output modality	Text
Precision	BF16
Reasoning effort	xhigh (default), medium, low
Tool calling	Yes
Streaming	Yes
Release date	October 2026

License

agnes-3.0-qwen is a post-trained derivative of Qwen/Qwen3.8-Flash-Next. Additional post-training was performed by Agnes AI. The model weights and inherited upstream artifacts are distributed under the Qwen Community License 1.0.

All use and redistribution must comply with the full terms in LICENSE. The original Qwen copyright and permission notices must be retained in all copies or substantial portions of the upstream software. See NOTICE for attribution and Agnes AI's post-training contribution.

Commercial use conditions (summary; the full license controls):

• If the licensee or any of its affiliates conducts a Model as a Service (MaaS) or AI Work Assistant business, a separate license from Qwen is required before using the model or its derivative works for any commercial purpose. The internal-use exception applies only when the software, its outputs, and its underlying model capabilities are not made available to any third party. See LICENSE for the definitions and exclusions for these businesses.
• If a commercial product or service using the model or its derivative works has more than 100,000,000 monthly active users or more than US$20,000,000 in monthly revenue (or the equivalent in other currencies), the respective model name must be prominently displayed on its user interface. This display condition is separate from the commercial licensing requirement above.

The API examples in this README do not grant additional commercial rights under the upstream license. Any separate commercial authorization must be obtained from Qwen; the contact listed in LICENSE is model-business@notice.qwencloud.com.

Hardware Requirements

The Quickstart launch command uses 8-GPU tensor parallelism. The checkpoint is a 119-shard BF16 package of about 360 GB; 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 500 GB 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

REASONING MODEL

agnes-3.0-qwen 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.

SGLang

python -m sglang.launch_server \
    --model-path Agnes-AI/agnes-3.0-qwen \
    --served-model-name agnes-3.0-qwen \
    --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

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-3.0-qwen",
    "messages": [
      {
        "role": "user",
        "content": "Review this API handler for security issues and provide a corrected version."
      }
    ],
    "temperature": 1.0,
    "max_tokens": 4000
  }'

Python

pip install openai

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-3.0-qwen",
    messages=[
        {
            "role": "user",
            "content": "Design a fault-tolerant event processing architecture.",
        }
    ],
    temperature=1.0,
    max_tokens=4000,
)

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

Reasoning Effort

The chat template accepts a reasoning_effort argument (xhigh by default, or medium / low) and an enable_thinking switch. With an OpenAI-compatible server these are passed through extra_body:

response = client.chat.completions.create(
    model="agnes-3.0-qwen",
    messages=[{"role": "user", "content": "Summarize the trade-offs of event sourcing in three bullets."}],
    extra_body={"chat_template_kwargs": {"reasoning_effort": "low"}},
)

Set {"enable_thinking": False} to disable extended reasoning entirely for latency-sensitive calls.

Image Understanding

response = client.chat.completions.create(
    model="agnes-3.0-qwen",
    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

curl https://apihub.agnes-ai.com/v1/responses \
  -H "Authorization: Bearer ${AGNES_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "agnes-3.0-qwen",
    "input": "Create a step-by-step migration plan from a monolith to services.",
    "max_output_tokens": 4000
  }'

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	4000 or higher

Raise max_tokens if a response stops early.

Model Capabilities

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

agnes-3.0-qwen is especially well suited to terminal and repository-level coding agents, tool-enabled customer workflows, technical research, document synthesis, and visual analysis 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

@misc{agnes30qwen2026,
  title        = {agnes-3.0-qwen},
  author       = {{Agnes AI}},
  year         = {2026},
  month        = oct,
  howpublished = {Open-weight model},
  url          = {https://huggingface.co/Agnes-AI/agnes-3.0-qwen}
}