Ant Group’s inclusionAI just dropped Ling 3.0 Flash, and the numbers tell a story worth paying attention to. The model scores 38 points on the Artificial Analysis Intelligence Index — a substantial leap over its predecessor and enough to tie Qwen3.6 27B while using far fewer active parameters. DeepSeek V4 Flash still holds the top spot at 52 points, so this isn’t a changing of the guard. But Artificial Analysis calls Ling 3.0 Flash the smartest open model under 124 billion total parameters. No smaller model touches its score.
Hallucination improvement is where things get genuinely interesting. On the AA Omniscience test, the hallucination rate fell from 97 percent to 44 percent compared to the previous version. That’s still not great in absolute terms — nearly half of responses can be wrong on that benchmark — but the trajectory matters. The model now refuses to answer questions it lacks reliable information on far more often, which suggests inclusionAI made a deliberate tradeoff: fewer confident-sounding fabrications, more honest silences. Anyone who has wrestled with smaller open models knows that’s a meaningful shift.
Pricing is the other headline. On a per-token basis, Ling 3.0 Flash undercuts every comparably capable model, including Qwen3.6 27B. It does consume more tokens on complex tasks, so the per-task math is less dramatic than the per-token chart suggests. Still, even accounting for that token inflation, it stays cheaper than Qwen3.6 27B. The model ships under an MIT license, with weights on Hugging Face and availability through the inclusionAI API and DeepInfra. That licensing choice matters — MIT means businesses can actually build on this without legal headaches, which has been a sticking point with other open releases.
The broader context: open models have been closing the gap with frontier systems all year, but the real competition has shifted from raw benchmark scores to cost-efficiency and reliability. Ling 3.0 Flash lands squarely in that second conversation. A 44 percent hallucination rate won’t win trust in regulated industries, yet the combination of MIT licensing, aggressive pricing, and competitive intelligence at this size class could push developers who’ve been defaulting to larger, pricier models to reconsider. Whether inclusionAI can keep this momentum against whatever DeepSeek ships next is the question that actually matters.
