Reportedly: Ant Group open-sources Ling-3.1-flash model with up to 1M context
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Ant Group open-sourced Ling-3.1-flash, a ~560B total parameter model with ~25B active parameters per token and up to 1M context. In Artificial Analysis public evaluation, it scored 41 intelligence index and ~210 tokens/sec output speed. Local deployment uses MXFP4 quantization for routed experts and block-FP8 for dense layers, reducing weight footprint from ~534GB to 292GB and Hopper GPU requirement from 8 to 4 cards, IT Home reported.
China context
- Original name
- 蚂蚁集团
- Outside China
- Open weights · ithome.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can access the open weights and deploy the model locally with reduced GPU requirements (4 Hopper cards instead of 8) using the provided quantization scheme.
- For investors
- Ant Group's open-source release of a large-context model signals its commitment to the open-weights ecosystem, potentially influencing enterprise adoption and competitive dynamics in the LLM market.
Translated from Chinese. Quotes and facts link to the original sources.
Ant Group announced the open-source release of Ling-3.1-flash, a large language model with approximately 560 billion total parameters, activating about 25 billion per token, and supporting up to 1 million tokens of context. The model was evaluated on Artificial Analysis, achieving an intelligence index of 41 and an output speed of around 210 tokens per second. For local deployment, the team applied MXFP4 quantization to routed experts and kept dense layers in block-FP8, cutting weight storage from about 534GB to 292GB. In corresponding Hopper GPU tests, serving deployment dropped from 8 cards to 4 cards, enabling teams with server-grade GPUs to run the model with fewer cards; actual requirements still vary with context length and concurrency.
The model uses a mixture-of-experts architecture with ~560B total parameters and ~25B active per token. Quantization strategy: MXFP4 for routed experts, block-FP8 for dense layers, reducing memory footprint from ~534GB to 292GB. This allows serving on 4 Hopper GPUs instead of 8 in tested configurations. Context window up to 1M tokens.
Developers outside China can now access and deploy a 1M-context open-weights model from Ant Group, reducing hardware requirements for large-context inference from 8 to 4 Hopper GPUs through quantization. This lowers the cost barrier for teams with server-grade GPUs to run a 560B-parameter model locally.
For enterprises and developers, the reduced GPU requirement (from 8 to 4 Hopper cards) and smaller weight footprint (292GB vs 534GB) can significantly lower infrastructure costs for serving a large-context model. The 1M context window enables processing of very long documents or codebases without chunking, potentially improving workflows in legal, research, and software development.
Next observable signals: whether Ant Group releases additional Ling-3.1 variants (e.g., larger or smaller models), whether the model appears on Hugging Face with open weights, and whether independent benchmarks confirm the reported intelligence index and throughput. Also check for developer downloads and community forks of the model repository, and any licensing terms that may restrict commercial use.