Ling-3.1-flash-1019: Ant Group releases 560-billion-parameter agent-focused model under MIT license
Ant Group's inclusionAI lab is open-sourcing the weights of a reworked version of Ling-3.1-flash, rated 44 by Artificial Analysis, and also publishing the base model for anyone who wants to train it further.
inclusionAI, the AI lab of Ant Group (Alipay’s parent company), open-sourced the weights of Ling-3.1-flash-1019 this Tuesday, October 27 — a reworked version of its Ling-3.1-flash model. The model is published on Hugging Face and ModelScope under the MIT license, one of the most permissive out there: commercial use, modification, and redistribution are all free. The official announcement dropped on X at 1:45 PM Paris time.
Ant is also releasing Ling-3.1-flash-base, the pre-post-training base model, for those who want to adapt it or train it on their own data.
A very large model, but few active parameters
According to its model card, Ling-3.1-flash-1019 is a mixture-of-experts (MoE) model with 560 billion parameters, of which only 25 billion are engaged per token: 8 out of 512 experts are activated at each step. Ant describes the architecture as “hybrid linear”: most layers use a linear attention mechanism (KDA), cheaper on long texts, and one layer in eight uses classic attention (Gated MLA). This is what allows it to claim a one-million-token context window.
The 25-billion active parameter figure shouldn’t mislead: you still need to load the entire model into memory. The weights, mostly in FP8, weigh in at around 570 GB, and the reference configuration provided with SGLang runs on eight H200 GPUs (or four GB300). This is a server model, not a workstation one.
What’s changed from the previous version
The “1019” designation refers to additional post-training, focused on reasoning and agents (multi-step tasks with tools). The developer claims clear gains: 1,524 Elo on AA-Briefcase (roughly 1,400 before), 68% on AutomationBench-AA (62%), 26% on CritPt (18%), and 57% on SciCode (54%).
The most solid number is the overall index: Artificial Analysis, which measures models independently, does show a score of 44 on its Intelligence Index, versus 41 for the previous version according to Ant. The site lists a price of $0.30 per million input tokens and $0.90 per million output tokens.
Limits to keep in mind
The same Artificial Analysis page flags one weakness: the model is very verbose. It generated 350 million tokens to get through the index’s test battery, versus a median of 140 million. In practice, that means longer, slower responses, and a bill that climbs faster than the per-token price suggests: Artificial Analysis puts the average cost per index task at $1.49.
The detailed agent scores (AA-Briefcase, AutomationBench) remain, at this stage, those communicated by the developer. Finally, Artificial Analysis dates the model’s release to October 22, while the public repo was only created on the 27th: it may have been evaluated or served via API before the weights were opened, without that being specified.
Our take
The main appeal isn’t the three-point gain, but the combination: a model this size, under MIT license, with its base model and a one-million-token context window, remains rare. Chinese labs keep publishing weights where major U.S. players mostly sell APIs. For a company, the practical question lies elsewhere: running 560 billion parameters requires heavy infrastructure, and the verbosity measured by Artificial Analysis can eat into part of the pricing advantage. Worth testing on your own tasks before seeing it as a replacement for closed models.
Sources
- Hugging Face — model card inclusionAI/Ling-3.1-flash-1019 Primary source
- Hugging Face — model card inclusionAI/Ling-3.1-flash-base Primary source
- Ant Ling (X) — open weights announcement Primary source
- Artificial Analysis — Ling 3.1 Flash 1019 page Secondary source — Independent measurement of Intelligence Index, pricing, and verbosity
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
