Ling 3.0 Flash sets open-weights efficiency frontier
Artificial Analysis scored Ant Group's 124B open-weights model at 38, placing its 5B active parameters ahead of every flash-tier rival.
The benchmark
Artificial Analysis published independent results for Ant Group's Ling 3.0 Flash late on Friday UTC, scoring the model 38 on version 4.1.1 of its Intelligence Index. The evaluator said the 124-billion-parameter mixture-of-experts model, which activates roughly 5 billion parameters per token, extends the open-weights Pareto frontier for intelligence against active parameters, sitting well ahead of flash-tier peers that activate 10 billion or more. It also placed the model on the frontier for intelligence against total parameters.
The sharpest movement was in hallucination behaviour. On AA-Omniscience — a metric that penalises confidently wrong answers rather than simply rewarding correct ones — Ling 3.0 Flash scored -18, a 48-point improvement over the -66 recorded by the previous-generation Ling 2.6 Flash. Artificial Analysis attributed the gain primarily to a lower hallucination rate rather than to broader knowledge coverage.
The model
Ling 3.0 Flash comes from inclusionAI, Ant Group's model lab. Weights were published earlier this month on Hugging Face and ModelScope under an MIT licence, in BF16 (roughly 255GB) and FP8 (roughly 128GB) variants, with deployment supported through SGLang and vLLM. The architecture uses a native hybrid-linear attention scheme that alternates Kimi Delta Attention and multi-head latent attention layers at a five-to-one ratio, supporting 256K context natively and extensible towards 1M. Ant has claimed the model matches or beats its own trillion-parameter flagship on most of the benchmarks it published.
Why it matters
Active-parameter count, not total size, drives serving cost. A model that answers at roughly 5 billion active parameters while scoring in the same band as far heavier open releases changes the arithmetic for high-volume agent deployments, where a single user request can trigger dozens of model calls. Independent confirmation matters more than the vendor's own charts, and the hallucination result is the operationally significant part: small-active models have historically bought their speed by guessing more, which is precisely what breaks automated pipelines. The result also underlines where Chinese labs are concentrating their effort — competing on the cost-per-unit-intelligence frontier rather than chasing absolute leaderboard peaks.