InclusionAI, the open-source group behind Alibaba affiliate Ant's Ling models, released Ling-3.0-flash, a freely downloadable reasoning model that uses only a small fraction of its parameters for each word it generates. It handles very long inputs and is aimed at coding and multi-step agent tasks, with reported scores at or above the group's far larger previous flagship. Running it locally still requires a four-GPU server, though it is also offered free through OpenRouter.
What changed
The group's previous flagship, Ring-2.6-1T, needed roughly eight times the total parameters to reach comparable benchmark scores.
What it unlocks
Running a long-context reasoning and agent model with open weights on a four-GPU server, or trying it free through OpenRouter.
- 124B total parameters, 5.1B active
- ~12.4% of the parameters of the 1T-class Ring-2.6-1T
- 256K context window
- Time to First Token reduced 60% to over 80% on long inputs
- SWE-Bench Pro 56.6, SWE-Bench Multilingual 72.4
- AIME 2026 93.2, HLE 22.7
What you need to act on it
- four high-memory GPUs (141GB-class, or eight 80GB cards) for self-hosting
- SGLang or the project's vLLM fork
- huggingface.co2026-08-07