AI news · October 16, 2026
InclusionAI releases Ling 3.1 Flash with 560B parameters and a 1M-token context
InclusionAI released Ling 3.1 Flash with 560B total and 25B active parameters under MIT, plus a 1M-token context, so teams can compare an open agent model with Muse Spark Open.
- total parameters
- 560B
- active parameters
- 25B
- context tokens
- 1M
InclusionAI lists Ling 3.1 Flash as a 560 billion parameter hybrid-linear MoE model with 25 billion active parameters and a one-million-token context. The model card uses an MIT license and describes agent, coding, office and finance use cases. It recommends SGLang for serving, lists day-one image support and says a full-context run needs eight H200 GPUs or four GB300 GPUs. The card also gives setup paths for Claude Code, Kilo, Qwen Code, Hermes and OpenClaw.
Ling is a different hardware trade from Meta's Muse Spark Open: it has a smaller total size and active count, while both target long-context agent work. The strong task claims are model-card claims, not independent results. The MIT license is easier for a commercial prototype, but the hardware and serving work still set a high cost. Measure useful output per dollar on the exact coding or document task before choosing it for production.
What you can do with it
Try the model card's hosted or local route if you have the listed hardware, then run a fixed coding task and a long document task. Compare answer quality, tool-call success, memory, tokens per accepted result and total serving cost with a smaller model before scaling.
Our take
Ling 3.1 Flash makes long-context open models more practical to compare, especially where the MIT license matters. It is still a serious machine purchase, so the right first move is a short hosted or shared-server trial. Do not confuse a large context window with reliable use of every token.
Links Try it
Source: InclusionAI / Hugging Face — Made With Models writes the brief; the reporting is theirs.