IT之家 · 智能时代 · 10/16/2026, 06:58:04
Ant Group Open-Sources Ling-3.1-flash: 560B MoE Model with 1M Context and Halved VRAM via Quantization
Ant Group has open-sourced Ling-3.1-flash, a 560B-parameter Mixture-of-Experts model activating 25B parameters per token, supporting up to 1M context length. By employing mixed quantization (MXFP4 for routed experts and block-FP8 for dense layers), the model's weight footprint is reduced from 534GB to 292GB, enabling deployment on 4 Hopper GPUs instead of 8. Benchmarks indicate an average time-to-first-token of 604ms in long-context scenarios, accompanied by detailed throughput and cost metrics.
SOURCE COVERAGEFull Original Report
On October 16, Ant Group’s Ling team announced the open-source release of the Ling-3.1-flash model. Ling-3.1-flash features approximately 560B total parameters, with about 25B activated per token, and supports a maximum context length of 1M tokens.
In Artificial Analysis’s public benchmarks, Ling-3.1-flash achieved an intelligence index of 41 and an output speed of approximately 210 tokens per second.
For local deployment, the team applied MXFP4 quantization to routed experts while maintaining block-FP8 for dense layers, reducing weight memory usage from approximately 534GB to 292GB. In corresponding Hopper GPU test configurations, service deployment requirements dropped from 8 GPUs to 4 GPUs, enabling teams with server-grade GPUs to run the model with fewer cards. Actual resource needs still vary depending on context length, concurrency levels, and service configuration.
To handle long-document processing, the team provides a parallel inference solution for long contexts (PCP / DCP). Tool calling has also been enhanced with improved JSON Schema support, facilitating integration with existing business interfaces.
Enterprise knowledge Q&A and office assistants often require reading lengthy documents first, yet users expect rapid responses. The team reduced inference overhead through operator fusion and DSpark speculative decoding: in simulated production requests averaging ~25.47K input tokens, ~224 output tokens, and a 76% cache hit rate, the average Time to First Token (TTFT) under an 8-GPU Hopper configuration was 604 milliseconds, with an average inter-token latency of 3.28 milliseconds. This provides a testing reference for interactive response times after ingesting long documents.
Business systems requiring continuous service must also evaluate how many requests a single instance can handle within a fixed GPU budget. In throughput tests using 4-GPU Hopper configurations with average inputs of ~8.5K tokens and outputs of ~800 tokens, the model achieved approximately 1.36 QPS, while meeting test constraints of TTFT P50 < 3 seconds and inter-token latency P90 < 70 milliseconds. This offers a reference for capacity planning in self-hosted deployments.
Regarding online access, Ling-3.1-flash currently offers a limited-time 25% discount. Post-discount pricing is as follows:
- OpenRouter (USD pricing): Input 0.225; Cache Read $0.015.
- Ling Studio (CNY pricing): Input ¥0.5; Output ¥1.5; Cache Read ¥0.1.
IT Home provides the open-source link below: