| --- |
| license: mit |
| --- |
| <p align="center"> |
| <img src="https://mdn.alipayobjects.com/huamei_qa8qxu/afts/img/A*4QxcQrBlTiAAAAAAQXAAAAgAemJ7AQ/original" width="100"/> |
| </p> |
| <p align="center">🤗 <a href="https://huggingface.co/inclusionAI">Hugging Face</a> | 🤖 <a href="https://modelscope.cn/organization/inclusionAI">ModelScope </a> | 🐙 <a href="https://openrouter.ai/inclusionai/ling-3.0-tiny:free">OpenRouter </a> </p> |
| |
| # Introduction |
| We are introducing **Ling-3.0-tiny**, a lightweight hybrid reasoning MoE model with **7.9B** total parameters and only **1.3B** activated parameters per token. |
| It is designed to deliver strong reasoning and agentic capabilities at low inference cost, making advanced model capabilities more accessible for local and resource-constrained deployment. |
| BF16, FP8, and INT4 weights are provided for a wide range of hardware and deployment settings. |
|
|
| Key highlights of the model are summarized below: |
|
|
| + **Efficient Hybrid-Linear Architecture:** Ling-3.0-tiny integrates a 3:1 alternating stacking of KDA and MLA (3 Kimi Delta Attention layers followed by 1 Multi-Head Latent Attention layer per 4-layer block) with a sparse MoE FFN comprising 128 routed experts. Each token activates only 8 routed experts and 1 shared expert, allowing the model to balance long-context modeling capability, parameter efficiency, and computational cost. |
| + **Native Hybrid Reasoning and Agentic Capabilities:** Ling-3.0-tiny supports both fast responses and multi-step reasoning, with thinking mode configurable per request through `enable_thinking`. It delivers balanced performance across general agent tasks, coding, mathematical and scientific reasoning, and instruction following. |
| + **Local and Edge Deployment:** Designed for efficient local deployment, Ling-3.0-tiny has been validated on **NVIDIA DGX Spark, Apple Silicon MacBook, and Mac mini**, enabling capable reasoning and agentic workloads without datacenter-class GPUs. With FP8, Ling-3.0-tiny reaches around **100-105 tokens/s on DGX Spark** and **86-90 tokens/s on an M4 Pro MacBook**, with approximately **8.34 GiB peak memory usage** at an 8K context length. |
|
|
| # Model Overview |
| Ling-3.0-tiny inherits the hybrid linear attation architecture of Ling-3.0 series, while being specifically optimized for lightweight and accessible deployment. The model has 7.9B total parameters, with only 1.3B parameters activated per token. |
|
|
| The architecture of Ling-3.0-tiny is designed to make computational efficiency serve real-world agentic performance. |
|
|
| + A 3:1 KDA–MLA architecture (3 KDA layers and 1 MLA layer per 4-layer block) provides more efficient long-context processing; |
| + A sparse MoE FFN with 128 experts activates 8 routed experts and 1 shared expert per token, enabling broad model capabilities with only 1.3B activated parameters per token. |
| + Native hybrid reasoning enables fast responses for routine tasks and multi-step reasoning for complex tasks within a single model. |
|
|
| Overall, these designs deliver the inference efficiency needed to deploy lightweight models in real-world agentic workflows. |
|
|
| <!-- Ling-3.0-tiny architecture diagram --> |
|  |
|
|
| # Evaluation |
| We evaluated Ling-3.0-tiny across agentic tasks, coding, long-context understanding, knowledge reliability, mathematical and scientific reasoning, and instruction following. |
| Ling-3.0-tiny achieves a score of **25** on the Artificial Analysis Intelligence Index v4.1.1 and **16** on the Artificial Analysis Agentic Index. |
| In Artificial Analysis testing, Ling-3.0-tiny reaches an output speed of over **160 tokens/s**, with approximately **18 seconds** of end-to-end latency for a 500-token response, including reasoning time. |
| These results highlight the model's efficiency relative to its 1.3B activated parameter footprint. |
|
|
| The following table presents representative benchmarks for Ling-3.0-tiny: |
|
|
|  |
|
|
| > + Thinking mode is enabled by default. The recommended sampling parameters for Ling-3.0-tiny are `temperature=1.0`, `top_p=0.95`, and `top_k=20`. |
| > + Terminal-Bench 2.1: Evaluated under the Artificial Analysis (AA) protocol using the default Terminus 2 harness, a unified 2-hour timeout, the provided JSON parser in preserve-thinking mode, and 3 runs per task (mean). Decoding uses temperature=1.0, max_new_tokens=32K, with a 256K context window. |
|
|
| # Quickstart |
| ## SGLang |
| The hardware- and recipe-specific launch matrix (BF16/FP8 × Low-Latency / High-Throughput / HiCache + Mooncake), with a live command generator and verified configurations, lives in the SGLang cookbook: |
|
|
| **Cookbook:** https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.0-tiny |
|
|
|
|
| ### Install SGLang |
| Use the pre-built image that tracks the Ling-3.0 runtime: |
|
|
| ```bash |
| docker pull lmsysorg/sglang:dev-Ling-3.0-tiny |
| ``` |
|
|
| ### Run Inference |
| Recommended low-latency recipe (built-in MTP / NEXTN, 256K YaRN context) on 1× 141GB-class GPU (H20-3e) or a 1-GPU Blackwell node: |
|
|
| **Server** |
|
|
| ```bash |
| docker run --rm --gpus all --ipc=host --shm-size 32g \ |
| -p 30000:30000 \ |
| -e HF_TOKEN=<your-hf-token> \ |
| lmsysorg/sglang:dev-Ling-3.0-tiny \ |
| env SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 \ |
| python3 -m sglang.launch_server \ |
| --model-path inclusionAI/Ling-3.0-tiny \ |
| --tp 1 \ |
| --json-model-override-args '{"rope_scaling":{"rope_type":"yarn","factor":2.0,"rope_theta":6000000,"partial_rotary_factor":0.5,"original_max_position_embeddings":131072}}' \ |
| --context-length 262144 \ |
| --speculative-algorithm NEXTN \ |
| --mem-fraction-static 0.8 \ |
| --host 0.0.0.0 \ |
| --port 30000 |
| ``` |
|
|
| **Client** |
|
|
| Thinking is enabled by default by both the chat template and the `ling3` reasoning parser. Disable it per request with `"chat_template_kwargs": {"enable_thinking": false}`. We recommend the sampling parameters `temperature=1.0`, `top_p=0.95`, and `top_k=20`. |
|
|
| ```bash |
| curl -s http://localhost:30000/v1/chat/completions \ |
| -H "Content-Type: application/json" \ |
| -d '{"model": "auto", |
| "messages": [{"role": "user", "content": "What is the capital of France?"}], |
| "stream": true, |
| "temperature": 1.0, |
| "top_k": 20, |
| "top_p": 0.95 |
| }' |
| ``` |
| For `--reasoning-parser ling3` / `--tool-call-parser ling3`, the HiCache + Mooncake L3 setup, and GSM8K / bench_serving reproduction commands, see the cookbook page linked above. |
| |
| ## vLLM |
| ### Install vLLM with Ling-3.0 Support |
| ```bash |
| pip install uv |
| |
| uv venv ~/my_ling_env |
| |
| source ~/my_ling_env/bin/activate |
| |
| git clone -b ling_3_0 https://github.com/inclusionAI/vllm-ling-v3.git |
| |
| cd vllm-ling-v3 |
| |
| VLLM_USE_PRECOMPILED=1 uv pip install --editable . --torch-backend=auto |
| ``` |
| |
| ### Run Inference |
| Here is the example to run Ling-3.0-tiny with a single GPU, where the server port is `${PORT}`: |
| |
| **Server** |
| |
| ```bash |
| vllm serve "$MODEL_PATH" \ |
| --port "$PORT" \ |
| --trust-remote-code \ |
| --served-model-name auto \ |
| --tensor-parallel-size 1 \ |
| --gpu-memory-utilization 0.85 \ |
| --enable-prefix-caching \ |
| --mamba-cache-mode align \ |
| --enable-auto-tool-choice \ |
| --tool-call-parser ling3 \ |
| --reasoning-parser ling3 |
| ``` |
| |
| **Client** |
|
|
| For better performance, We recommend setting `enable_thinking=true` with `temperature=1.0`, `top_p=0.95`, and `top_k=20`. |
|
|
| ```bash |
| curl -s http://${MASTER_IP}:${PORT}/v1/chat/completions \ |
| -H "Content-Type: application/json" \ |
| -d '{"model": "auto", |
| "messages": [{"role": "user", "content": "What is the capital of France?"}], |
| "chat_template_kwargs": {"enable_thinking": true}, |
| "stream": true, |
| "temperature": 1.0, |
| "top_k": 20, |
| "top_p": 0.95 |
| }' |
| ``` |
|
|
| ## Ollama |
| > + This configuration has been verified on an M4 Pro Mac with 48 GB of unified memory. |
|
|
| ### Preparation and Build |
| ```bash |
| git clone https://github.com/ollama/ollama.git |
| cd ollama |
| git fetch origin refs/pull/17643/head:bailing-moe-v3 |
| git switch bailing-moe-v3 |
| |
| cmake -B build . |
| cmake --build build --parallel 8 |
| ``` |
|
|
| > + Support is currently provided by [ollama/ollama#17643](https://github.com/ollama/ollama/pull/17643) and is limited to running via MLX on Apple Silicon. |
| > + Use the local `./ollama` executable built from source in this section. This functionality is not yet included in the official Ollama release. |
|
|
| ### Import Model |
| Replace `/absolute/path/to/bf16_weights` with the absolute path to the BF16 model weights directory. The imported model will be named `ling-tiny-bf16`: |
|
|
| ```bash |
| printf 'FROM /absolute/path/to/bf16_weights\n' > /tmp/Modelfile.ling |
| ./ollama create ling-tiny-bf16 --experimental -f /tmp/Modelfile.ling |
| ``` |
|
|
| ### Start Service |
| Set the default context length to 8192, and then start the Ollama service: |
|
|
| ```bash |
| # The service listens on http://127.0.0.1:11434 by default |
| OLLAMA_CONTEXT_LENGTH=8192 ./ollama serve |
| ``` |
|
|
| ### Call API |
| ```bash |
| curl -sS http://127.0.0.1:11434/api/generate -d '{ |
| "model": "ling-tiny-bf16", |
| "prompt": "<role>SYSTEM</role>detailed thinking on<|role_end|><role>HUMAN</role>Calculate 17 × 23 and output only the number.<|role_end|><role>ASSISTANT</role>\n<think>", |
| "raw": true, |
| "think": true, |
| "stream": false, |
| "options": { |
| "temperature": 1.0, |
| "top_p": 0.95, |
| "top_k": 20, |
| "num_predict": 2048 |
| } |
| }' | jq -r .response |
| ``` |
|
|
