Ling 3.1 Flash

Architectural Overview & Capability Profile

Ling 3.1 Flash (released as inclusionAI/Ling-3.1-flash-1019 on Hugging Face) is InclusionAI’s flagship open-weights foundation model. Featuring an advanced 560 billion total parameter Mixture-of-Experts (MoE) architecture with 25 billion active parameters per token, Ling 3.1 Flash combines hybrid linear attention mechanisms with multi-step reinforcement learning from agent environments. On RankLLMs’s calibrated evaluation framework, Ling 3.1 Flash earns an overall benchmark rating of 57.4 / 100, placing it among the premier open-weights developer models globally.

Distributed under a permissive MIT Open-Weights License, Ling 3.1 Flash natively pairs with LingCode—an open agent harness built upon Pi that provides tool orchestration, context preservation, and iterative result verification across an expansive 1,048,576-token context window.

Official Hugging Face Release Benchmarks

Verified multi-suite evaluations released on Hugging Face demonstrate Ling 3.1 Flash trading blows with proprietary frontier flagships like Claude Opus 5 and GPT-5.6 Sol:

1. Coding & Repository Engineering:
   • FrontierSWE: Scores 75.16%, leading GLM-5.3 (74.34%), Kimi K3 (73.52%), and GPT-5.6 Sol (71.30%).
   • SWE-bench Pro: Reaches 68.74%, surpassing GPT-5.6 Sol (64.60%), GLM-5.3 (61.83%), and DeepSeek-V4.1-Flash (56.77%).
   • SWE-bench Multilingual: Scores 88.50% across cross-lingual codebases.
   • Terminal-Bench 4.0: Achieves 46.03% in the LingCode harness, outperforming GLM-5.3 (41.90%), GPT-5.6 Sol (39.90%), and DeepSeek-V4.1-Flash (31.20%).
   • SWE Atlas - Codebase QnA: Achieves 64.51%, leading GPT-5.6 Sol (54.00%) and GLM-5.3 (59.00%).
2. Defensive Cybersecurity & Health:
   • CyberGym: Achieves 87.90%, capturing the top spot across all open and closed frontier models evaluated (surpassing GPT-5.6 Sol at 84.50% and GLM-5.3 at 84.50%).
   • HealthBench Professional: Scores 65.35% in the AQ environment, leading GPT-5.6 Sol (60.76%), Claude Opus 5 (59.80%), and DeepSeek (50.37%).
3. Agentic Workflows & Multi-Challenge:
   • BrowseComp: Scores 91.67%, surpassing Claude Opus 5 (90.80%) and GPT-5.6 Sol (90.40%).
   • DRACO: Reaches 85.49%, outscoring DeepSeek (79.85%) and GPT-5.6 Sol (77.70%).
   • GDPval-AA v2.1: Records 1,673 Elo, outperforming DeepSeek-V4.1-Flash (1,600 Elo) and GPT-5.6 Sol (1,588 Elo).
   • AutomationBench (Public): Scores 52.50%, beating Claude Opus 5 (50.30%) and GLM-5.3 (48.20%).
   • MultiChallenge: Delivers 69.78%, outpacing GPT-5.6 Sol (66.32%) and Claude Opus 5 (64.98%).

Economics & Inference Efficiency

Ling 3.1 Flash delivers exceptional throughput at low cost:

• Streaming Speed: 165 tokens per second at peak throughput.
• API Pricing: $0.20 per 1M input tokens and $0.60 per 1M output tokens (blended ~$0.30/1M tokens).
• Deployment: Full FP8/4-bit weights are available on Hugging Face for direct deployment on vLLM and SGLang clusters.