On August 4, Ant Group's inclusionAI team pushed the weights of Ling-3.0-Flash to Hugging Face (repo: inclusionAI/Ling-3.0-flash). The release cadence was tight: OpenRouter listing on July 23, official announcement July 24, free API ending August 3 at 23:00, and open weights on August 4.
Key specs
- Total parameters: 124B; activated parameters: 5.1B
- Compared to Ant's own 1T-parameter flagship thinking model Ring-2.6-1T, it is 12.4% of total parameters and 8.1% of activated parameters
- Official claim: per-token compute cost ≈ 1/12 of the flagship
- Deployable on a single 8-GPU node, versus dozens of GPUs typically required for trillion-parameter models
- MiniAppBench (multi-step task execution): 25.3% pass rate vs. a 17% average across 16 mainstream models
- As an execution node, 42.9% better performance than the flagship's own execution layer
- Qwen3 series: closed flagship + open mid-size models; generalist, multi-size coverage
- DeepSeek-V3: open-source, ultra-large scale, low-cost training; relatively higher per-token activation
- Llama 4 (Meta): multimodal, multilingual, large-activation generalist route
- Ling-3.0-Flash: extreme sparsity + native hybrid linear attention + explicit Agent-execution-node positioning — a narrower, sharper niche
- Independent benchmarks missing: Artificial Analysis had no independent evaluation page as of July 24; all figures (MiniAppBench 25.3%, TTFT reduction, 1/12 cost) are official claims.
- License: Ant claims Apache 2.0 for Ling-3.0-Flash (Ling-2.6-flash used MIT), pending confirmation on the model card.
- No public data yet on SWE-bench, HumanEval, or LiveCodeBench — the benchmarks that matter most if it truly aims to be an Agent execution node.
- Whether the KDA + MLA 5:1 + 1/64 MoE recipe can be reproduced and extended by other teams will shape Ant's influence on Chinese LLM architecture directions.
- Hugging Face: https://huggingface.co/inclusionAI/Ling-3.0-flash
- Ant Ling official announcement: https://mp.weixin.qq.com/s?__biz=MzkyODk2MDQwNw%3D%3D&mid=2247487457&idx=1&sn=24ad4a355d81291e53fbe680ca987112
- OpenRouter: https://openrouter.ai/models/inclusionai/ling-3.0-flash:free
- Aliyun developer analysis: https://developer.aliyun.com/article/1751301
- TMTPost: https://www.tmtpost.com/agent/ai-article/19326
- Digital Applied: https://www.digitalapplied.com/blog/ling-3-0-flash-ant-group-efficiency-moe
Architecture highlights
1. Native hybrid linear attention from pretraining: KDA (Kimi Delta Attention, adopted from Moonshot AI's Kimi Linear research) and MLA (Multi-head Latent Attention, DeepSeek-style) stacked alternately at a 5:1 ratio, plus a new fine-grained diagonal gating for KDA. 2. 1/64 sparsity MoE — double the previous generation's 1/32 — further compressing per-token activated experts. Combining attention slimming and expert-routing slimming is relatively rare among Chinese open-source MoE models, where most teams optimize only one axis.
Positioned as an "Agent execution node"
Rather than chasing general-purpose capability, Ling-3.0-Flash targets the execution layer in a planning-execution split: complex tasks go to a 1T-class thinking model (Ring-2.6-1T), while Ling-3.0-Flash handles high-frequency mid-workflow execution. Official numbers:
End-to-end open-source inference stack
Ant integrated SGLang HiCache and Mooncake tiered KV caching into long-context inference, cutting first-token latency (TTFT) by 60–80%+ on long inputs. The result is a fully Chinese-led stack: model weights (Hugging Face) + inference framework (SGLang, LMSYS-led) + tiered caching (Mooncake, from Moonshot AI) — rather than "Chinese model + foreign inference stack."
OpenRouter production-grade measurements: 0.81s end-to-end latency, 3 tokens/s throughput, 99.96% 30-day availability. The free-access strategy follows Ling-2.6-flash (104B/7.4B activated, MIT-licensed, open-sourced in April) and Ling-2.6-1T — the entire Ling family has been free on OpenRouter since Ling 2.0 (October 2025).
Comparison with peer MoE models
Open questions
Bottom line
Ling-3.0-Flash is not just another Chinese open-source MoE. It is the first clear attempt to (a) position a model explicitly as an Agent-workflow execution node, and (b) assemble a complete Chinese-led open inference chain: weights + SGLang + Mooncake. If independent evaluations confirm the official claims, the cost curve for high-concurrency Agent workloads in 2026 H2 may be materially rewritten.