license: mit
🤗 Hugging Face | 🤖 ModelScope | 🐙 OpenRouter
Introduction
We have open-sourced the Ling-3.0 series, our most efficient language foundation model family to date. To support research and community-driven innovation, we are releasing a collection of checkpoints during the training process as following:
| Model | Pre-trained | Mid-trained | Merged (i.e., WSM) |
|---|---|---|---|
| Ling-3.0-tiny | Ling-3.0-tiny-base-30T | Ling-3.0-tiny-base-midtrain | Ling-3.0-tiny-base |
| Ling-3.0-flash | Ling-3.0-flash-base-30T | Ling-3.0-flash-base-midtrain | Ling-3.0-flash-base |
These checkpoints correspond to different stages of the training process:
- Pretrained checkpoint have completed large-scale pretraining but have not undergone mid-training, WSM merging (or learning-rate decay), or post-training.
- Mid-trained checkpoint have completed mid-training but have not undergone WSM merging (or learning-rate decay) or post-training.
- Merged checkpoints have undergone WSM merging (or learning-rate decay) based on the mid-training checkpoints but have not undergone post-training.
These checkpoints are released to support continued pretraining, fine-tuning, and further research. For the post-trained model, please see Ling-3.0-tiny and Ling-3.0-flash.
Model Overview
Key features
- Highly sparse (1/64) MoE architecture: 512 routed experts, with only 8 routed experts and 1 shared expert activated per token. This enables broad model capabilities while activating just 5.1B (Non-emb) parameters per token;
- Native hybrid linear attention: Ling-3.0 series adopt a native hybrid linear attention architecture from the very start of pretraining by combining KDA with Gated MLA to enable efficient processing of long-context inputs.
- Warmup-Stable and Merge: We replace conventional learning-rate decay with weighted checkpoint merging. By eliminating the decay phase, our Base Model is better suited for continual pretraining and dynamic data expansion, while enabling offline exploration of different decay profiles without rerunning costly experiments for each strategy.
- Scale Seamlessly: Ling-3.0-tiny-base and Ling-3.0-flash-base share the same training recipe, enabling community to experiment on the Ling-3.0-tiny-base first and then scale validated training strategies to the larger Ling-3.0-flash-base.
| Model Type | Base (final checkpoint of WSM merging) |
|---|---|
| Architecture | Hybrid-linear MoE |
| Parameter Scale | Totoal 7.9B, Activated 1.3B |
| Transformer Layers | 18 KDA + 6 Gated MLA (3:1) |
| Number of Dense Layers | 1 |
| Number of Routed Experts | 128 |
| Number of Shared Experts | 1 |
| Number of Activated Experts | 8 |
| Attention Heads | 16 |
| Hidden Size | 1536 |
| Expert Intermediate Size | 512 |
| Dense Intermediate Size | 4608 |
| Vocabulary Size | 157,184 |
Base Model Evaluation
To systematically assess the capabilities of the base model, we use a self-built comprehensive benchmark suite covering several key domains, including knowledge, coding, mathematics, reasoning, multilingual understanding, and long-context comprehension. The performance of the pretrained base checkpoint, i.e., Ling-3.0-tiny-base, is compared below:
| Domain | Benchmark | Shot Config | Ling-3.0-tiny-base | Ling-2.5-mini-base | Qwen3.5-9B-base | Qwen3.5-4B-base |
|---|---|---|---|---|---|---|
| 7.9B A1.3B | 16B A1.4B | 9B | 4B | |||
| Knowledge | CCPM(EM) | 0-shot | 84.01 | 80.77 | 88.31 | 81.88 |
| ARC-C(EM) | 0-shot | 92.20 | 91.19 | 94.58 | 92.20 | |
| AGIEval(Acc) | 0-shot | 64.49 | 63.38 | 66.17 | 61.45 | |
| SimpleQA-Verified(Acc) | 5-shot | 7.20 | 7.10 | 10.40 | 6.40 | |
| MMLU-Pro(EM) | 5-shot | 51.83 | 49.89 | 57.76 | 52.06 | |
| CEval(EM) | 5-shot | 80.59 | 80.16 | 81.52 | 76.67 | |
| Code | HumanEval-Plus(Pass@1) | 0-shot | 79.27 | 76.22 | 52.44 | 49.39 |
| CruxEval(Pass@1) | 1-shot | 67.44 | 63.19 | 68.12 | 64.31 | |
| MultiPL-E(Pass@1) | 1-shot | 64.38 | 64.34 | 52.89 | 45.47 | |
| LiveCodeBench1(Pass@1) | 1-shot | 24.23 | 22.91 | 18.06 | 13.88 | |
| BigCodeBench(Pass@1) | 0-shot | 42.89 | 42.54 | 29.39 | 24.39 | |
| FullStackBench(Pass@1) | 3-shot | 39.48 | 38.59 | 37.23 | 33.14 | |
| LCBench2(Pass@1) | 3-shot | 41.59 | 46.16 | 33.47 | 19.46 | |
| Math | MATH500(Acc) | 4-shot | 65.60 | 68.40 | 55.20 | 49.60 |
| OlympiadBench(Acc) | 3-shot | 25.90 | 24.10 | 21.84 | 17.02 | |
| TheoremQA(Acc) | 5-shot | 51.64 | 50.68 | 52.88 | 47.40 | |
| OmniMath(Acc) | 3-shot | 29.70 | 29.02 | 22.99 | 20.89 | |
| Reasoning | CommonSenseQA(EM) | 5-shot | 83.46 | 81.16 | 83.62 | 80.18 |
| BBH(EM) | 3-shot | 80.64 | 77.37 | 84.52 | 80.11 | |
| Long-context | LongBench(Acc) | 0-shot | 43.85 | 29.93 | 51.87 | 39.96 |
| LEval(Acc) | 0-shot | 68.37 | 62.72 | 77.24 | 63.09 | |
|
Note: 1 LiveCodeBench (2408-2505) 2 LCBench (2301-2502) |
||||||
Intended Use
Recommended use cases:
- Continued pre-training
- Mid-training
- Supervised fine-tuning for domain adaptation
- Preference optimization and RL post-training Distillation research
- Long-context and MoE systems research
Not recommended as-is for:
- Direct end-user chat deployment
- Safety-critical applications without additional alignment and evaluation
- Production use without post-training and task-specific validation
Usage
For fine-tuning examples, please refer to our ling-cookbook.
FAQ
If you have any question, please feel free to add a discussion.
License
This model is released under the MIT License.