Ling 3.0 Flash gives the open model market a new reference point for capability at a smaller scale. Based on the Artificial Analysis Intelligence Index, the model reaches 38 points, placing it level with Qwen3.6 27B while using far fewer active parameters.
That combination matters because open models are often judged not only by how smart they are, but by how much compute they require, how reliably they answer, and how practical they are to deploy. Ling 3.0 Flash stands out on all three fronts in the source data, even though it does not lead every ranking.
A higher score in a smaller class
Artificial Analysis places Ling 3.0 Flash at 38 points on its Intelligence Index. The score represents a major improvement over the model's predecessor and gives it the same index result as Qwen3.6 27B.
The more important comparison is not only the score, but the model size context around it. Ling 3.0 Flash reaches that level while using far fewer active parameters, which makes the result notable for teams and researchers watching the tradeoff between model capability and runtime footprint.
The model still has a clear gap to the top performer named in the source. DeepSeek V4 Flash sits at 52 points, keeping it ahead of Ling 3.0 Flash on the same broad comparison.
Even with that gap, Ling 3.0 Flash has a specific claim that sets it apart: according to Artificial Analysis, it is the smartest open model under 124 billion total parameters. The source also states that no smaller model matches its score.
Hallucination performance improves sharply
The most striking reliability change appears in the AA Omniscience test. Compared with the previous version, Ling 3.0 Flash lowered its hallucination rate from 97 to 44 percent.
That shift is important because a model's usefulness depends on more than producing fluent output. A model that gives an answer when it lacks reliable information can create a different kind of failure than one that admits uncertainty.
The source says Ling 3.0 Flash now refuses to answer questions it does not have reliable answers for far more often. In practical terms, that means the model is not only being measured on whether it can answer, but also on whether it can avoid answering when the answer would be unsupported.
This is a meaningful distinction for open models. Better refusal behavior can make a model more predictable in workflows where unsupported answers are costly, even if the model is not perfect and still has a measured hallucination rate.
Agentic task gains add another layer
Ling 3.0 Flash also improves over its predecessor on agentic tasks. The source specifically mentions gains on the t3-Bench Banking benchmark.
Agentic tasks are different from simple question answering because they usually require a model to follow steps, make decisions across a task, or handle structured interactions. The source does not provide the exact t3-Bench Banking score, so the safest conclusion is limited: Ling 3.0 Flash is reported to be stronger than the previous version on this benchmark.
That matters because modern model comparisons increasingly look beyond static knowledge tests. A model can be valuable if it performs well in task-oriented settings, especially when those gains come alongside stronger refusal behavior and a competitive intelligence score.
Pricing and availability strengthen the case
Cost is another point where Ling 3.0 Flash is presented as unusually competitive. On per-token pricing, it beats every comparably capable model named by the source.
There is one caveat. Ling 3.0 Flash uses more tokens on complex tasks than similarly strong alternatives. That means per-token pricing alone does not tell the full cost story, because a model that consumes more tokens can narrow the pricing advantage when measured by completed task.
Even with that caveat, the source says Ling 3.0 Flash remains cheaper than Qwen3.6 27B on a per-task basis. That gives the model a stronger practical position: it is not only inexpensive by token, but still cheaper after accounting for higher token use on complex work.
Ant Group's inclusionAI is releasing Ling 3.0 Flash under an MIT license. The model is available through the inclusionAI API and DeepInfra, while model weights are available directly on Hugging Face.
Those release details matter because they define how the model can be accessed. API availability makes testing and hosted use possible, while direct model weights on Hugging Face give users another path to work with the model under the stated license.
What the benchmark picture says
The source data presents Ling 3.0 Flash as a model with a clear profile: strong for its size, substantially improved over its predecessor, and cost-competitive against similarly capable options.
It does not displace DeepSeek V4 Flash at the top of the cited Intelligence Index comparison. But it does create a different kind of headline by reaching 38 points under 124 billion total parameters and matching Qwen3.6 27B while using far fewer active parameters.
For the open model field, that is the core takeaway. Ling 3.0 Flash is not simply another release with a higher benchmark number. It combines a stronger intelligence score, a lower AA Omniscience hallucination rate, reported agentic improvements, lower pricing, and broad release channels into one package.