Ling 3.0 AI model, represented by a glowing brain, cuts hallucination rate in half for open models.

Editorial illustration for Ling 3.0 Flash Cuts Hallucination Rate in Half for Open Models

Ling 3.0 Flash Cuts Hallucination Rate in Half

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

August 13, 2026 • 4 min read

Ant Group's inclusionAI has released Ling 3.0 Flash, and the numbers put it ahead of every open model under 124 billion total parameters. On the Artificial Analysis Intelligence Index, the model scores 38 points, a sharp climb from its predecessor and enough to match Qwen3.6 27B despite running with far fewer active parameters. DeepSeek V4 Flash still leads the pack at 52 points, but among smaller open models, nothing currently beats Ling 3.0 Flash's score.

The bigger story might be reliability. Open models have a well-documented tendency to answer confidently when they should just say "I don't know," and Ling 3.0 Flash appears to have made real progress there. The model also picked up gains on agentic benchmarks, including t3-Bench Banking, and undercuts comparable models on price per token, even though it tends to use more tokens on harder problems.

Ant Group is releasing the model under an MIT license, with access through the inclusionAI API, DeepInfra, and weights posted directly on Hugging Face. Here's what changed on the hallucination front:

Ling 3.0 Flash is the smartest open model in its size class. On the Artificial Analysis Intelligence Index, Ling 3.0 Flash scores 38 points, a big jump over its predecessor. That puts it on par with Qwen3.6 27B while using far fewer active parameters.

Ling 3.0 Flash is the smartest open model at its size - THE DECODER

Why this matters

The 44-percent hallucination rate is still a coin-flip's worth of wrong answers on Omniscience, so calling this "solved" would be premature. What's more interesting than the topline intelligence score is the behavior shift: a model that's learned to say "I don't know" instead of confabulating is more useful in production than one that scores a few points higher but bluffs constantly. For developers picking open weights for RAG pipelines or agent backends, that refusal behavior often matters more than raw benchmark rank, because a wrong answer delivered confidently costs more to catch than an honest non-answer.

Landing between Qwen3.6 27B and DeepSeek V4 Flash on Artificial Analysis's index, while using fewer active parameters, also says something about where efficiency gains are actually coming from this cycle, calibration and refusal training, not just scale. We'd want to see how Ling 3.0 Flash performs on domain-specific hallucination tests before trusting it in anything regulated. Worth watching whether "smartest under 124B" holds once the next Qwen or DeepSeek refresh lands.

Common Questions Answered

How does Ling 3.0 Flash's hallucination rate compare to previous versions?

Ling 3.0 Flash has cut the hallucination rate in half compared to its predecessor, achieving a 44-percent hallucination rate on the Omniscience benchmark. While this represents significant improvement, the article notes this is still a substantial error rate and should not be considered a complete solution to the hallucination problem.

What is Ling 3.0 Flash's score on the Artificial Analysis Intelligence Index?

Ling 3.0 Flash scores 38 points on the Artificial Analysis Intelligence Index, which is a sharp climb from its predecessor and puts it on par with Qwen3.6 27B despite using far fewer active parameters. This score makes it the highest-performing open model under 124 billion total parameters, though DeepSeek V4 Flash still leads overall at 52 points.

Why is Ling 3.0 Flash's refusal behavior more valuable than raw intelligence scores?

A model that learns to say 'I don't know' instead of confabulating is more useful in production environments than one that scores higher but provides incorrect confident answers. For developers building RAG pipelines or agent backends with open weights models, this behavioral shift toward honesty about knowledge limitations is more practical than marginal improvements in benchmark scores.

Which company released Ling 3.0 Flash and what parameters does it operate with?

Ant Group's inclusionAI released Ling 3.0 Flash, and the model operates with far fewer active parameters than competitors like Qwen3.6 27B while achieving comparable performance scores. This efficiency makes it particularly competitive in the open model space under 124 billion total parameters.

Ship an AI product this weekend — no engineers required.

Structured, in-depth lessons on the exact no-code tools — not scattered tutorials.

The exact platforms, taught in depth
Build real, working projects
Our honest review + a reader discount

Read the review →