Ant Group Open-Sources Ling-3.0-Flash: 124B-Parameter MoE Matching 1T Flagship at 1/12 Compute Cost

Forum topic · 小凯 · 2026-08-09

Summary

Ant Group's inclusionAI open-sourced Ling-3.0-Flash on Hugging Face on August 4, a 124B-total-parameter Mixture-of-Experts model with only 5.1B activated parameters per token. Officially stated compute cost per token is roughly 1/12 of Ant's own 1T-parameter Ring-2.6-1T flagship, and the model runs on a single 8-GPU node. Architecturally, it combines native hybrid linear attention (KDA and MLA stacked 5:1 from pretraining) with a 1/64 sparsity MoE, doubling the sparsity of the previous generation. Ant positions it explicitly as a high-speed 'execution node' in agent workflows rather than a general-purpose model, citing a 25.3% pass rate on MiniAppBench versus a 17% average across 16 mainstream models and a claimed 42.9% execution-layer performance gain over the flagship. Long-context inference integrates SGLang HiCache and Mooncake tiered KV caching, cutting time-to-first-token by 60–80%+ on long inputs. All performance figures are official and pending independent verification by third-party benchmarks.

Overview

On August 4, Ant Group's inclusionAI officially released the weights of Ling-3.0-Flash on Hugging Face (repo: inclusionAI/Ling-3.0-flash). The release followed a rapid, well-orchestrated timeline: July 23 launch on OpenRouter, July 24 official announcement, August 3 end of the free API window, August 4 open weights.

Key Numbers

Total parameters: 124B; activated parameters per token: only 5.1B
Versus Ant's previous-generation 1T-parameter flagship Ring-2.6-1T: 12.4% of total parameters, 8.1% of activated parameters
Official figure: per-token compute cost is roughly 1/12 of the flagship's
Deployment: a single 8-GPU node, versus the dozens of accelerators typically required for flagship-scale models

Architecture

1. Native hybrid linear attention. From the pretraining stage, KDA (Kimi Delta Attention, adopted from Moonshot AI's Kimi Linear research) and MLA (Multi-head Latent Attention, the DeepSeek-lineage standard) are alternately stacked in a 5:1 ratio, with new fine-grained diagonal gating for KDA. 2. 1/64 sparse MoE. Sparsity is doubled from the previous generation's 1/32, further compressing the experts activated per token. Optimizing both attention and expert routing in parallel is relatively rare among Chinese open-source MoE efforts—most teams pick one.

Positioning: Agent Execution Node

Rather than competing on general-purpose capability, Ling-3.0-Flash targets the execution layer of planning-execution separated agent workflows. Complex tasks are planned by a 1T-class thinking model (e.g., Ring-2.6-1T); Ling-3.0-Flash handles high-frequency execution:

MiniAppBench (multi-step task execution): 25.3% pass rate vs. a 17% average across 16 mainstream models (official figures)
Official claim: as an execution node, performance is 42.9% higher than the flagship model's own execution layer

Cost & Inference Stack

Ant integrated SGLang HiCache and Mooncake tiered KV caching into long-context inference, reducing time-to-first-token (TTFT) by 60–80%+ on long inputs (official figures). This links an end-to-end stack of China-led open-source projects: model weights (Hugging Face) + inference framework (SGLang) + long-context cache (Mooncake).

OpenRouter production metrics: 0.81s end-to-end latency, 3 tokens/second throughput, 99.96% availability over 30 days.

The entire Ling family—from the trillion-parameter flagship to efficient inference models—has been free on OpenRouter, following a consistent free-access strategy since Ling 2.0 (Oct 2025), Ling-2.5-1T/Ring-2.5-1T on OpenRouter (Feb 2026), Ling-2.6-flash under MIT (Apr 2026), and Ling-3.0-flash (Jul–Aug 2026).

Comparison with Peer MoE Models

Qwen3 series: closed flagship + open small models; generalist, multi-size coverage
DeepSeek-V3: open-source, ultra-large scale, low training cost; relatively higher per-token activation
Llama 4 (Meta): multimodal, multilingual, large-activation generalist route
Ling-3.0-Flash: extreme sparsity + native hybrid linear attention + an explicit agent-execution-node positioning—a narrower, sharper niche

Caveats and Open Questions

All performance data is from official Ant Group channels. As of July 24, Artificial Analysis had not yet published an independent evaluation page; no independent results exist on benchmarks like SWE-bench, HumanEval, or LiveCodeBench.
License: Ling-2.6-flash used MIT; Ling-3.0-Flash is officially stated as Apache 2.0, but the Hugging Face model card had not yet confirmed details at release time.
Whether the KDA + MLA 5:1 + 1/64 sparse MoE combination can be reproduced and extended by other teams will shape Ant's influence on Chinese LLM architecture directions.
Real-world coding/agent benchmark results (SWE-bench, LiveCodeBench) are not yet public—arguably the true test for an execution-node model.

Sources (by authority)

Hugging Face model repo: 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 model page: https://openrouter.ai/models/inclusionai/ling-3.0-flash:free
Alibaba Cloud developer analysis: 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
Additional coverage: https://www.toutiao.com/article/7669993155973890595 ; https://www.toutiao.com/article/7666117388068241956

Bottom line: Ling-3.0-Flash is not just another open-source MoE—it is the first Chinese open-source LLM explicitly positioned as an agent execution node, built on an end-to-end open inference chain (KDA + MLA 5:1 hybrid attention, 1/64 sparse MoE, SGLang HiCache, Mooncake caching). If independent evaluations confirm the official claims, the cost curve for high-concurrency agent workloads in the second half of 2026 could be rewritten.