inclusionAI/Ling-3.0-tiny-int4

🤗 Hugging Face   |   🤖 ModelScope    |   🐙 OpenRouter   

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 under a small inference compute footprint, 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 uses a 3:1 alternating stacking of KDA and MLA (3 Kimi Delta Attention layers followed by 1 Multi-Head Latent Attention layer in every 4-layer block), combined with a sparse MoE 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, tool use, 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 not designed to optimize a single technical metric; instead, it aims to translate efficiency into real-world task performance.

• A 3:1 KDA–MLA architecture (3 KDA layers and 1 MLA layer in every 4-layer block) enables more efficient long-context processing;
• A sparse MoE 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.

Together, these designs target three goals:

• Reduce the computational resources required for inference;
• Lower the barriers to local deployment and downstream development;
• Enable lightweight models to participate in real-world agent workflows.

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:

• All Ling-3.0-tiny evaluation results use Thinking mode; the recommended sampling parameters 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:

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

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.

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.