AI BriefinginclusionAI releases Ling-3.1-flash: A 560B-parameter hybrid reasoning model with 1M-token context
Key point
The model features 25B active parameters per token and demonstrates strong performance on agentic benchmarks like Terminal-Bench 4.0 and SWE-bench Pro.
Details
inclusionAI has introduced Ling-3.1-flash, an upgraded native hybrid reasoning model designed for complex real-world tasks. The model utilizes a Hybrid-linear MoE architecture with 560B total parameters and 25B active parameters per token, supporting a 1M-token context window. It builds upon Ling-3.0-flash with architectural improvements aimed at enhancing performance in general-purpose agents, search, office productivity, and software engineering.
Architecture and Specifications
The model's structure consists of 63 KDA and 9 Gated MLA layers in a 7:1 ratio, along with 4 dense layers. Key specifications include:
- Routed Experts: 512
- Shared Experts: 1
- Activated Experts: 8
- Attention Heads: 64
- Hidden Size: 4096
- Vocabulary Size: 157184
Performance and Benchmarks
Ling-3.1-flash shows improvements across various domains, including healthcare, finance, and materials science. It performs strongly on representative coding and agentic benchmarks such as Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE, GDPVal-AA V2.1, Automationbench, MCP-Atlas, and SkillsBench. The model is also compatible with frameworks like Claude Code, Kilo Code, Qwen Code, Hermes Agent, and OpenClaw.
Additionally, inclusionAI introduced LingCode, a harness built on the open-source coding agent Pi. LingCode is enhanced to support long-horizon coding tasks involving hundreds of tool calls and repeated context compaction, delivering significant performance gains on specific coding agent tasks.
Deployment and Availability
The model is available on Hugging Face, ModelScope, and OpenRouter. For inference, inclusionAI recommends using SGLang with day-0 support, including MTP (NEXTN) speculative decoding and Hopper decode optimizations. Verified recipes are provided for hardware configurations including H200, GB300, H800, H100, and B200. vLLM support is also available, requiring specific environment preparation and configuration flags for optimal performance.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.