LaunchAI ModelsAugust 2, 2026

inclusionAI releases Ling-3.0-flash, a 127.5B-parameter model

127.5B total params, 5.1B active, 512 experts with 8 active per token; MIT license with BF16 (~255GB) and official FP8 (~128GB) weights on Hugging Face. Reddit testers report ~80 tok/s decoding on a single DGX Spark.

How this story unfolded

3 weeks · 2 reports · 12 community posts · 14 of 15 shown

Jul 23
- Ling 3.0 Flash is now available on AI Gatewayvercel.com
Jul 25
- inclusionAI (Ant Group) released Ling-3.0-flash, a 124B sparse MoE (5.1B active), 256K context, API-only on OpenRouter and free for a week
Jul 27
- Ling-3.0-flash weights: SGLang says day-0, vLLM says when they land, llama.cpp closed the 2.6 request as not_planned
Jul 31
- Review testing on Ling 3.0 flash - From one prompt to a 3D world
Aug 4
- inclusionAI/Ling-3.0-flashhuggingface.co
Aug 5
- Ling-3.0-flash MXFP4 released and running locally on one DGX Spark.
Aug 8
- Ant Group has just released Ling 3.0 Flash, a 124B open weights model that scores 38 on the...
Aug 9
- Two flags took the official Ling-3.0-flash INT4 from 20.8 to 38.7 tok/s on one DGX Spark
Aug 11

Daily brief

Get tomorrow's AI brief in your inbox

inclusionAI releases Ling-3.0-flash, a 127.5B-parameter model

How this story unfolded

Daily brief

More stories today