이미지: X — 벤치마크·평가 화면 갈무리
Summary
- Ant Group has released Ling 3.0 Tiny, an open-weight small model that scored 25 on the Artificial Analysis Intelligence Index.
- With 7.9B total parameters and 1.3B active parameters, it sits on the Pareto frontier for "active parameters versus intelligence."
- Its accuracy was just 9%, the lowest among the models compared, but because it only attempted 37% of questions, its hallucination rate came in at 30%, the lowest in the group — at the cost of the highest output-token usage per task, at 51k.
- 모델명
- Ling 3.0 Tiny (앤트그룹, 오픈웨이트)
- 파라미터
- 총 7.9B / 활성 1.3B
- Intelligence Index
- 25점 (Artificial Analysis v4.1.1)
- AA-Omniscience Index
- -19 (Nemotron 3.5 Lightning -18에 이어 표 내 2위)
- 정답률
- 9% (비교 목록 최하위, 최고는 Muse Glimmer high 27%)
- 환각률
- 30% (비교 목록 최저, Ministral 3 8B는 94%)
- 시도율
- 37% — 질문의 3분의 1가량에만 답을 시도
- 과제당 출력 토큰
- 51k (Ling 3.0 Flash 36k, Nemotron 3.5 Lightning·Qwen3.6 35B A3B는 각 31k)
If it doesn't know, it doesn't answer
There's usually one way for a small model to beat a big one: don't pretend to know. Ant Group's open-weight small model, Ling 3.0 Tiny, only attempted answers for 37% of knowledge questions. It skipped the rest. As a result, its rate of giving wrong answers — its hallucination rate — came in at 30%, the lowest among the 12 models compared.
According to measurements by independent evaluator Artificial Analysis, the model scored 25 on the Intelligence Index. Its total parameter count is 7.9B, but only 1.3B active parameters actually activate when processing a single query. Artificial Analysis noted that the model sits on the Pareto frontier for "active parameters versus intelligence" — meaning no smarter model exists yet at this size.
How small is 1.3B active parameters, really
Mixture-of-Experts (MoE) architectures don't activate the entire model — only the parts needed for a given task. That's why total parameter counts can be large while active parameters stay small. By comparison, NVIDIA's Nemotron 3.5 Lightning has 30B total parameters with 3B active, while Ant Group's flagship model Ling 3.0 Flash, released August 7–8, has 124B total. Ling 3.0 Tiny's 1.3B stands out as notably small even within this category. Smaller models consume less memory, and lower memory demands make it feasible to run them on a laptop or a single GPU.
Last in accuracy, high on the index — how?
The AA-Omniscience Index measures the reliability of a model's knowledge. Correct answers earn points, wrong answers lose points, and admitting uncertainty costs nothing. The scale runs from -100 to 100: zero means correct and incorrect answers are equal in number, and negative scores mean wrong answers outnumber right ones. Small models, which simply hold less knowledge to begin with, tend to land in negative territory.
| Model | Omniscience Index | Accuracy | Hallucination Rate |
|---|---|---|---|
| Nemotron 3.5 Lightning | -18 | 14% | 38% |
| Ling 3.0 Tiny | -19 | 9% | 30% |
| Qwen3.6 35B A3B | -22 | 19% | 51% |
| Muse Glimmer (high) | -33 | 27% | 82% |
| Ministral 3 8B | -69 | 13% | 94% |
Read the table row by row and the pattern becomes clear. Muse Glimmer has the highest accuracy at 27%, but also an 82% hallucination rate — because it attempts to answer even what it doesn't know. Ling 3.0 Tiny has the lowest accuracy on the list at 9%, yet ranks second on the index at -19. Artificial Analysis attributed this to its "low attempt rate reducing hallucinations."
In practice, this distinction matters. For tasks like document summarization or internal search, where a single wrong fact can ruin an entire output, a model that stays silent when uncertain is easier to work with than one that gets half its answers right and fabricates the rest. Conversely, it's not well suited to chatbots that need broad, general-knowledge responses.
The cost is paid in tokens
Nothing is free. Ling 3.0 Tiny used an average of 51k output tokens to solve a single Intelligence Index task. Of that, 32k were reasoning tokens spent working through the problem before producing an answer.
| Model | Output tokens per task | |
|---|---|---|
| Ling 3.0 Tiny | 51k | 100 |
| Ling 3.0 Flash | 36k | 71 |
| Nemotron 3.5 Lightning | 31k | 61 |
| Qwen3.6 35B A3B | 31k | 61 |
| Gemma 4 31B | 5k | 10 |
According to Artificial Analysis figures, that's roughly 65% more than Nemotron 3.5 Lightning or Qwen3.6 35B A3B. It effectively compensates for fewer parameters with longer chains of reasoning. Because the model is small, it can generate each token quickly — but if it needs to generate far more tokens overall, that speed advantage in latency and cost is offset. This is why parameter count alone isn't enough when choosing a small reasoning model.
So what does this change
A wave of small open-weight models has arrived in recent weeks. NVIDIA's Nemotron 3.5 Lightning scored 24 on the Intelligence Index, Ling 3.0 Flash scored 38, and now Ling 3.0 Tiny adds a 25. On score alone, they're all roughly comparable. What differs is character. Some models compete on speed, others on breadth of knowledge — Ling 3.0 Tiny has staked its position on "saying less that's wrong."
Being open-weight, its parameters can be downloaded and run independently. For use cases like internal document Q&A, where giving a wrong answer is riskier than giving no answer at all, this adds one more option worth testing on your own servers. Still, whether a low attempt rate on benchmarks translates into an overuse of "I don't know" in actual production use is something you'd need to test firsthand. It's not something a single metric can settle.
