Ant Group open-sources Ling-3.0-flash: 124B params, only 5.1B active

Ant Group's inclusionAI open-sourced the Ling-3.0-flash model on Hugging Face on August 4. The model has approximately 124B total parameters but activates only 5.1B per token, using a MoE architecture, supporting 256K context, and released under the pure MIT license. The official claims its performance rivals flagship models with 1T activated parameters, and the FP8 quantized version requires only 128GB VRAM for deployment. It is now available on the AI/ML API platform with a free trial period. This combination of extremely low activation and high cost-effectiveness has attracted widespread community attention.

Confirmed

• Architecture and parameters: Total parameters range from 124B to 127.5B (slight variations across posts), with 5.1B activated parameters (about 1/64 of total). It uses a MoE architecture with 512 routed experts and 1 shared expert, and natively incorporates a hybrid linear architecture from the pretraining stage, supporting 256K context.
• Deployment cost: Official BF16 and FP8 quantized versions are provided; the FP8 version requires only 128GB VRAM, significantly lowering the hardware barrier for such a large model.
• Features and benchmark performance: The model supports thinking mode and tool calling, and provides complete chat template configuration. Official reports show strong performance on benchmarks like SWE-bench, and the official claims it can rival flagship models with 1T activated parameters.
• Open source and API availability: The model is released under the pure MIT license with no commercial restrictions. It is also available on the AI/ML API platform with a free trial period.

Why it matters

• Ling-3.0-flash balances model capacity and inference cost through extremely low parameter activation, significantly reducing deployment barriers while maintaining high performance, offering the open-source community a highly cost-effective large-scale hybrid inference model.

2026-08-04 ~ 2026-08-06 · 11 related posts

Full story(4 episodes)→

• Episode 1: Ant Group Releases Ling-3.0-flash MoE Model(2026-07-24, 12 posts)
• Episode 2: Ant Group Releases Ling-3.0-flash Hybrid Reasoning Model(2026-07-27, 2 posts)
• Episode 3: Ant Group Releases Ling-3.0-flash Hybrid Reasoning Model(2026-07-29, 2 posts)
• Episode 4: Ant Group open-sources Ling-3.0-flash: 124B params, only 5.1B active(2026-08-04, 11 posts)

• Ling-3.0-flash Open-Sourced: 124B Hybrid-linear MoE Model Released — NielsRogge · 2026-08-04
• [source] Ling-3.0-flash Open-Sourced: 127B MoE with Official FP8 Requiring Only 128GB — derspenti · 2026-08-04
• Ling-3.0-flash Open Weights Released with BF16 and FP8 Versions — aftahi_ai · 2026-08-05
• Ant's Ling-3.0-flash Activates Only 5.1B Params: Architecture and Cost Analysis — jkris050 · 2026-08-05
• [source] Ant's Ling-3.0-flash Open Sourced: 124B Total, 5.1B Active with Impressive Benchmarks — Asleep-Pilot-4142 · 2026-08-05
• Ling 3.0 Flash Launches: 5.1B Active Params Matches 1T Flagship — AdinaYakup · 2026-08-05
• inclusionAI Releases Ling 3.0 flash Hybrid-Linear Reasoning Model — AdinaYakup · 2026-08-05
• Ant Ling 3.0 Flash Open-Weighted: 124B Params Rivals 1T Flagship — FellMentKE · 2026-08-05
• inclusionAI Releases Open Weights for Ling-3.0-flash Model — FellMentKE · 2026-08-05
• Ant's Ling-3.0-flash Hits API: 124B Params for Low-Cost Agents — rohanpaul_ai · 2026-08-06
• [source] Ant Group Open-Sources 124B Model Under Pure MIT License with No Restrictions — Asleep-Pilot-4142 · 2026-08-06