Marco-Mini-Instruct
Marco-Mini-Instruct is the instruction-tuned variant of Marco-Mini-Base, a highly sparse Mixture-of-Experts (MoE) multilingual language model from the Marco-MoE family, developed by Alibaba International Digital Commerce. It activates only 0.86B out of 17.3B total parameters (5% activation ratio) per token. Marco-Mini-Instruct achieves the best average performance across English, multilingual general, and multilingual cultural benchmarks when compared against instruct models with up to 12B activated parameters, including Qwen3-4B-Instruct, Ministral3-8B-Instruct, Gemma3-12B-Instruct, LFM2-24B-A2B, and Granite4-Small-Instruct.
Model Description
Marco-Mini-Instruct shares the same architecture as Marco-Mini-Base: a decoder-only Transformer with sparse MoE layers replacing standard FFN layers, upcycled from Qwen3-0.6B-Base using fine-grained sub-matrix splitting combined with Drop-Upcycling.
| Configuration | Value |
|---|---|
| Total Parameters | 17.3B |
| Activated Parameters | 0.86B |
| Activation Ratio | 5% |
| Num Layers | 28 |
| Model Dimension | 1024 |
| FFN Intermediate Dimension | 3072 |
| Q-Heads | 16 |
| KV-Heads | 8 |
| Head Dimension | 128 |
| Expert Dimension | 768 |
| Total Experts | 256 |
| Activated Experts | 8 |
| Tie Embeddings | True |
| Training FLOPs | $1.56 \times 10^{23}$ |
Post-Training Details
Marco-Mini-Instruct is trained from Marco-Mini-Base using a two-stage post-training pipeline implemented with the SLIME framework:
Stage 1: Supervised Fine-Tuning (SFT)
- Duration: ~24 hours on 64 GPUs
- Steps: ~4,000 (1 epoch)
- Learning rate: 1e-5 with cosine decay to 1e-6
- Batch size: 512, context length 8,192 tokens
Data sources:
- General instructions — Dolci-Instruct dataset, augmented with Nemotron-Cascade-2 data
- Knowledge-intensive data — Scientific prompts from Nemotron-Cascade-2, responses distilled from Gemini3-Flash
- Translation data — Web-mined NLLB translation pairs, filtered and scored with Qwen3-Embedding-8B (top 10K per language)
- Multilingual & cultural data — Wikidata-sourced content with Gemini3-Flash text synthesis for cultural concepts.
Stage 2: On-Policy Distillation (OPD)
- Duration: ~110 hours on 64 GPUs
- Steps: ~3,800 total (2 responses sampled per prompt)
- Learning rate: 1e-6 (constant)
Cascaded distillation:
- ~1,900 steps with Qwen3-30B-A3B-Instruct as teacher
- ~1,900 steps with Qwen3-Next-80B-A3B-Instruct as stronger teacher
OPD data mixture:
| Category | Datasets | Ratio |
|---|---|---|
| Instruction Following | Nemotron-RL-instruction-following + structured outputs | 25% |
| Knowledge & Reasoning | Nemotron-RL-ReasoningGym-v1 + knowledge-mcqa | 25% |
| Alignment | Nemotron-Cascade-RL-RLHF | 10% |
| Math | DAPO-Math-17k + Skywork-OR1-RL-Data | 10% |
| Multilingual | Translation + Cultural + Nemotron-SFT-Multilingual-v1 | 30% |
Supported Languages
English, Chinese, Arabic, German, Spanish, French, Korean, Japanese, Portuguese, Turkish, Indonesian, Italian, Dutch, Polish, Russian, Vietnamese, Thai, Hebrew, Ukrainian, Malay, Bengali, Czech, Urdu, Kazakh, Greek, Romanian, Hungarian, Nepali, Azerbaijani
Evaluation
We compare Marco-Mini-Instruct against strong instruct baselines: Qwen3-4B-Instruct (4B activated), Ministral3-8B-Instruct (8.8B activated), Gemma3-12B-Instruct (12B activated), Granite4-Small-Instruct (9B activated), and LFM2-24B-A2B (2B activated). Marco-Mini-Instruct uses only 0.86B activated parameters. Avg@8 accuracies are reported, except for GlobalMMLU and MMMLU where Acc@1 is reported.
English
| Benchmark | Qwen3-4B | Ministral3-8B | Gemma3-12B | Granite4-Small | LFM2-24B-A2B | Marco-Mini |
|---|---|---|---|---|---|---|
| MMLU (Acc) | 80.8 | 79.8 | 76.2 | 76.7 | 74.9 | 83.4 |
| MMLU-Redux (Acc) | 80.9 | 79.9 | 76.2 | 76.7 | 74.9 | 83.5 |
| MMLU-Pro (Acc) | 66.9 | 63.9 | 55.8 | 57.1 | 57.6 | 70.7 |
| AGIEval (Acc) | 51.7 | 52.4 | 43.6 | 44.7 | 49.0 | 55.4 |
| GPQA-Diamond (Acc) | 50.8 | 44.8 | 35.2 | 38.6 | 39.7 | 50.3 |
| GSM8K (EM) | 88.6 | 89.5 | 89.7 | 83.9 | 87.2 | 93.1 |
| MATH (EM) | 93.4 | 86.2 | 83.8 | 75.7 | 83.9 | 91.8 |
| Average | 73.3 | 70.9 | 65.8 | 64.8 | 66.7 | 75.5 |
Multilingual — General
| Benchmark | Qwen3-4B | Ministral3-8B | Gemma3-12B | Granite4-Small | LFM2-24B-A2B | Marco-Mini |
|---|---|---|---|---|---|---|
| GlobalMMLU (Acc) | 70.2 | 55.4 | 69.2 | 67.4 | 57.0 | 73.3 |
| MMMLU (Acc) | 71.3 | 56.4 | 69.4 | 68.1 | 62.3 | 73.7 |
| MMLU-ProX-Lite (Acc) | 58.3 | 43.3 | 51.3 | 51.6 | 43.3 | 61.2 |
| MGPQA (Acc) | 41.0 | 30.5 | 32.8 | 35.0 | 32.7 | 41.8 |
| FLORES-200 En→Xx (BLEU) | 22.1 | 17.5 | 35.6 | 31.9 | 19.2 | 30.6 |
| FLORES-200 Xx→En (BLEU) | 33.5 | 31.0 | 40.3 | 32.2 | 22.7 | 36.8 |
| WMT24++ En→Xx (BLEU) | 20.9 | 14.4 | 32.1 | 26.6 | 16.0 | 26.8 |
| WMT24++ Xx→En (BLEU) | 29.9 | 24.2 | 35.5 | 27.5 | 18.8 | 31.3 |
| MGSM (EM) | 84.4 | 68.7 | 84.0 | 75.7 | 67.8 | 87.4 |
| PolyMath (EM) | 47.2 | 26.4 | 35.5 | 28.9 | 29.3 | 44.7 |
| Average | 47.9 | 36.8 | 48.6 | 44.5 | 36.9 | 50.8 |
Multilingual — Cultural & Regional
| Benchmark | Qwen3-4B | Ministral3-8B | Gemma3-12B | Granite4-Small | LFM2-24B-A2B | Marco-Mini |
|---|---|---|---|---|---|---|
| INCLUDE (Acc) | 63.8 | 50.7 | 65.0 | 60.3 | 49.1 | 65.6 |
| Global-PIQA (Acc) | 79.6 | 61.3 | 82.2 | 80.2 | 69.0 | 84.2 |
| CMMLU (Acc) | 78.6 | 67.4 | 60.8 | 59.6 | 56.7 | 75.3 |
| C-Eval (Acc) | 80.4 | 68.0 | 59.7 | 59.4 | 56.7 | 75.4 |
| ArabicMMLU (Acc) | 66.0 | 41.4 | 70.1 | 66.3 | 61.3 | 67.8 |
| TurkishMMLU (Acc) | 71.6 | 48.2 | 64.4 | 57.9 | 33.4 | 74.7 |
| GreekMMLU (Acc) | 68.6 | 49.5 | 77.7 | 71.7 | 44.7 | 72.5 |
| KazakhMMLU (Acc) | 66.6 | 59.1 | 66.8 | 63.5 | 47.6 | 68.8 |
| IndoMMLU (Acc) | 64.4 | 52.4 | 65.3 | 59.6 | 42.7 | 65.7 |
| IndoCareer (Acc) | 62.2 | 53.4 | 63.2 | 56.3 | 43.7 | 64.4 |
| IndoCulture (Acc) | 58.7 | 47.8 | 69.6 | 59.3 | 44.2 | 67.1 |
| Average | 69.1 | 54.5 | 67.7 | 63.1 | 49.9 | 71.0 |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "AIDC-AI/Marco-Mini-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
messages = [
{"role": "user", "content": "What is the capital of France?"}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))
Note: vLLM is the recommended engine for deployment, as SGLang currently lacks support for MoE models with tied embeddings (see PR #20127). If SGLang is required for your workflow, please use the specific build at commit e5f48b32abff027d859a43b7d5ba3aece04471c7.
Citation
@article{marco-moe,
title={Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient Upcycling},
author={Fan Jiang, Yu Zhao, Chenyang Lyu, Tianqi Shi, Yichao Du, Feihu Jiang, Longyue Wang and Weihua Luo},
year={2026}
}
License
This model is released under the Apache 2.0 License.