Ant Group Ling Model Open Weights Enable 560B MoE Deployment on Four GPUs

Ant Group’s inclusionAI released the Ling-3.1-Flash model with open weights under the MIT license in mid-October 2026. The 560 billion parameter mixture-of-experts system targets long-horizon agent tasks and software development workflows.

Ant Group Ling Model Architecture Supports Efficient Inference

The model uses a hybrid-linear design with 72 layers that combine 63 key-dimension-aware linear attention blocks and 9 gated multi-head latent attention layers in a 7:1 ratio. It activates eight routed experts out of 512 plus one shared expert per token, resulting in roughly 25 billion active parameters. Quantization with MXFP4 on routed experts and block-FP8 on dense layers reduces the footprint from 534 GB to 292 GB, allowing reference inference on four Hopper GPUs.

Official documentation lists support for up to 1M token context, though current hosted endpoints expose 256K to 262K tokens with a 32K output limit in several configurations. The architecture includes a hidden size of 4,096, 64 attention heads and a vocabulary of 157,184 tokens. Day-one integration exists for SGLang and vLLM with MTP speculative decoding.

Compared with the earlier Ling-3.0-flash variant of 124 billion total parameters, the new release doubles reported scores on agent benchmarks such as SWE-Pro and Terminal-Bench. Vendor data place the Artificial Analysis Intelligence Index at approximately 41, with competitive results on HealthBench Professional and CyberGym.

MIT License and API Access Expand Commercial Options

Weights appeared on Hugging Face under the inclusionAI/Ling-3.1-flash repository and on ModelScope shortly after a two-week free API trial concluded around October 13–14. The MIT license permits unrestricted commercial use and redistribution with only attribution requirements. Hosted inference remains available through OpenRouter and Novita alongside the official Ant Group service.

Developers can now run the model locally for tasks that require extended context, such as repository-scale code analysis or multi-step tool orchestration. Early community efforts already target llama.cpp and Ollama ports, while further quantizations are expected to reduce hardware thresholds further. No large independent safety or benchmark suites have been published as of October 17, 2026.

The release forms part of Ant Group’s broader Ling series focused on MoE designs validated on domestic hardware. Organizations evaluating self-hosted agent platforms can now test the model without API rate limits or data egress constraints, subject only to standard MIT compliance.

Practical Steps for Teams Evaluating the New Weights

Teams should first verify hardware compatibility using the published 4-GPU reference configuration and test context scaling from 256K upward. Integration with existing agent frameworks can start via the SGLang or vLLM endpoints already supported. Organizations handling sensitive data may prefer local deployment over third-party gateways to maintain full control of prompts and outputs.

Future updates are expected to include expanded context on public APIs and additional community tooling. Monitoring release notes from inclusionAI and Hugging Face will help teams track quantization improvements and framework optimizations as adoption grows.