Agents A1: 35B MoE Long-Horizon Agentic Model

Agents A1 is a 35-billion parameter Mixture-of-Experts (MoE) agentic model developed by the InternScience group at the Shanghai Artificial Intelligence Laboratory. Built upon the Qwen3.5-35B-A3B base architecture, Agents A1 is designed to scale horizon capability—the length, depth, and coherence of multi-step task execution—activating only ~3 billion parameters per token for ultra-fast local inference.

🔬 Architecture & Training

• Mixture-of-Experts (MoE): 35B total parameters with ~3B active parameters per token, providing high computational efficiency.
• 256K Context Window: Supports up to 262,144 tokens, allowing the model to analyze large code repositories, extensive scientific literature, and long multi-turn interactions.
• Three-Stage Training Recipe:
  1. Full-Domain SFT: Foundation alignment for generic agentic behavior and tool calling.
  2. Domain-Level Teacher Training: Deep domain specialization across 6 distinct fields.
  3. Multi-Teacher Domain-Routed On-Policy Distillation: Distills knowledge from domain experts into a unified MoE routing structure.
• Long-Horizon Trajectory Execution: Optimized for generating and maintaining long-horizon knowledge-action trajectories averaging 45,000 tokens in length.

Full-Domain SFT ───► Domain-Level Teacher Training ───► Multi-Teacher Domain-Routed On-Policy Distillation ───► Agents A1 MoE

📊 Benchmarks & Performance

🚀 Quickstart & Serving with vLLM

vllm serve InternScience/Agents-A1 \
 --port 8000 \
 --tensor-parallel-size 1 \
 --max-model-len 262144 \
 --reasoning-parser qwen3 \
 --enable-auto-tool-choice \
 --tool-call-parser qwen3_coder

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="none"
)

response = client.chat.completions.create(
    model="InternScience/Agents-A1",
    messages=[
        {"role": "user", "content": "Analyze the codebase and construct a long-horizon plan for refactoring."}
    ],
    temperature=0.7,
)

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

🔗 Official Paper & Resources

• Official Website
• arXiv Paper (arXiv:2606.30616)
• Paper PDF Download
• GitHub Repository
• Hugging Face Model Collection