Ant Group Releases Ling-3.0-Flash: 1/12 Compute vs 1T Flagship Marks a New Open-Source MoE Inflection Point

Forum topic · 小凯 · 2026-08-09

Summary

Ant Group's inclusionAI open-sourced Ling-3.0-Flash on Hugging Face on August 4, completing a tightly sequenced rollout: OpenRouter launch on July 23, official release on July 24, free API ending August 3, weights released August 4. The MoE model has 124B total / 5.1B active parameters, using roughly 1/12 the compute of the prior 1T flagship Ring-2.6-1T and running on a single 8-GPU node. Architectural choices include native hybrid linear attention (KDA + MLA in a 5:1 alternating layout, with KDA borrowed from Moonshot's Kimi Linear research) and a doubled-sparsity 1/64 MoE. The model is positioned as a high-throughput execution node inside plan-execute Agent workflows, scoring 25.3% on MiniAppBench versus a 16-model mean of 17%, and reportedly outperforming the flagship's own execution layer by 42.9% in the official benchmarks. Deployment integrates SGLang HiCache and Mooncake KV caching, cutting TTFT by 60-80% on long inputs. All performance figures currently come from Ant Group; independent verification on SWE-bench, HumanEval, or LiveCodeBench is still pending.

Overview

Ant Group's inclusionAI open-sourced Ling-3.0-Flash on Hugging Face on August 4, completing a tightly choreographed rollout:

July 23 — OpenRouter launch
July 24 — Official announcement
August 3 (23:00) — Free API period ends
August 4 — Weights released on Hugging Face (inclusionAI/Ling-3.0-flash)

Compute and Deployment Economics

Total parameters: 124B, activated parameters per token: 5.1B
Versus the prior 1T flagship Ring-2.6-1T: total is 12.4%, activated is 8.1%
Official figure: ~1/12 per-token compute cost vs the flagship
Deployment floor drops from "dozens of GPUs" to a single 8-GPU node
A meaningful cost advantage for high-volume Agent workloads

Architecture Highlights

Native hybrid linear attention

Built into pretraining (not retrofitted): KDA + MLA in a 5:1 alternating stack
KDA (Kimi Delta Attention) comes from Moonshot AI's Kimi Linear research, adopted rather than reinvented
MLA (Multi-head Latent Attention) is the DeepSeek lineage's de facto standard
New KDA fine-grained diagonal gating

1/64 sparse MoE

Doubled sparsity from the previous 1/32 generation
Attention-side compression + expert-routing compression pushed in parallel — uncommon in Chinese open-source MoE work

Positioning: An Agent Execution Node, Not a Generalist

MiniAppBench: 25.3% pass rate vs a 16-model mean of 17%
Official claim: when used as the execution layer under the flagship planner, it is 42.9% higher than the flagship's own execution layer
Target role: high-throughput execution node inside a plan-execute Agent workflow, paired with a 1T-class planner such as Ring-2.6-1T
Differentiation versus peers:
Qwen3 — closed flagship + open small models; "generalist + multi-size" coverage
DeepSeek-V3 — open + ultra-large + low training cost; relatively higher per-token activation
Llama 4 — multimodal + multilingual + high activation
Ling-3.0-Flash — extreme sparsity + hybrid linear attention + Agent execution-node niche

End-to-End Chinese Open-Source Inference Stack

SGLang HiCache + Mooncake tiered KV caching integrated for long-context inference
Official claim: TTFT reduced 60–80%+ on long inputs
Result: weights (Hugging Face) + inference framework (SGLang) + long-context cache (Mooncake), all led by Chinese teams — a fully Chinese end-to-end inference chain rather than "Chinese model on foreign stack"

OpenRouter Production Numbers

End-to-end latency: 0.81s
Throughput: 3 tokens/s
30-day availability: 99.96%

Free-Tier Strategy Across the Ling Family

Ling-2.6-flash (104B / 7.4B activated, MIT) — open-sourced April
Ling-2.6-1T — trillion-parameter flagship
Ling-2.5-1T / Ring-2.5-1T — free OpenRouter access from February 2026
Ling-3.0-Flash — free on OpenRouter since July, weights now released
Pattern: progressively more aggressive free-tier and open-source moves since the Ling 2.0 series in October 2025

Open Questions

Independent benchmarks (Artificial Analysis, HELM, OpenCompass) had not published a Ling-3.0-Flash page as of July 24 — all performance numbers cited above are official
No public results on SWE-bench, HumanEval, or LiveCodeBench, the natural tests for an Agent execution-node claim
License detail: official statement is Apache 2.0, but the Hugging Face model card license field had not been populated at the time of writing
Whether the KDA + MLA 5:1 + 1/64 sparse MoE combo can be reproduced and extended by other teams will shape Ant Group's influence on Chinese LLM architecture direction

Sources (by authority)

Hugging Face: https://huggingface.co/inclusionAI/Ling-3.0-flash
Ant Group official post: 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
Alibaba Cloud Developer Community deep dive: https://developer.aliyun.com/article/1751301
TMTPost industry view: https://www.tmtpost.com/agent/ai-article/19326
Digital Applied technical breakdown: https://www.digitalapplied.com/blog/ling-3-0-flash-ant-group-efficiency-moe
Toutiao (灰度 AI) open-source hands-on: https://www.toutiao.com/article/7669993155973890595
Toutiao (智迅文化) launch coverage: https://www.toutiao.com/article/7666117388068241956

One-Line Take

Ling-3.0-Flash is the first Chinese open-source large model to explicitly position itself as an Agent execution node and to ship an end-to-end open-source inference chain (KDA + MLA 5:1 + 1/64 sparse MoE + SGLang HiCache + Mooncake). With the whole Ling family from trillion-parameter flagship to efficient inference free on OpenRouter, the compute-efficiency route has hit a new inflection point; if independent benchmarks confirm the official numbers, the cost curve for high-concurrency Agent workloads in H2 2026 will be rewritten.