🤗 Hugging Face | 🤖 ModelScope
引言
Ling-2.5-1T，普惠智能，即刻见效。
思考型模型抬高了智能的上限，而即时型模型则通过平衡效率与性能来拓展其覆盖范围——让 AGI 不仅更强大，也更易获取。作为 Ling 家族最新的旗舰即时型模型，Ling-2.5-1T 在模型架构、token 效率和偏好对齐方面实现了全面升级，旨在将普惠 AI 的质量提升到全新水平。
- Ling-2.5-1T 拥有 1T 总参数量（激活参数 63B）。相比上一代，其预训练语料从 20T tokens 扩展至 29T tokens。借助高效的混合线性注意力架构和精细化的数据策略，该模型在处理高达 1M tokens 的上下文长度时，仍能实现极高的吞吐量。
- 通过引入结合“正确性”与“过程冗余”的复合奖励机制，Ling-2.5-1T 进一步推动了即时型模型在效率与性能平衡方面的前沿。在相当的 token 效率水平下，Ling-2.5-1T 的推理能力显著超越其前代，接近通常消耗约 4 倍输出 tokens 的前沿“思考型模型”的水平。
- 通过精细化的对齐策略——例如双向 RL 反馈和基于 Agent 的指令约束验证——Ling-2.5-1T 在偏好对齐任务上相较上一代取得了显著提升，包括创意写作和指令遵循。
- Ling-2.5-1T 在大规模高保真交互环境中采用 Agentic RL 训练，兼容 Claude Code、OpenCode 和 OpenClaw 等主流 agent 平台。它在通用工具调用基准 BFCL-V4 上取得了领先的开源性能。
模型下载
| 模型 | 上下文长度 | 下载 |
|---|---|---|
| Ling-2.5-1T | 256K -> 1M (YaRN) | 🤗 HuggingFace 🤖 ModelScope |
注意：如果你对之前的版本感兴趣，请访问 Huggingface 或 ModelScope 中过往的模型合集。
部署
SGLang
环境准备
我们稍后会将模型提交至 SGLang 官方发布，现在我们可以按照以下步骤准备环境：
git clone -b ling_2_5 git@github.com:antgroup/sglang.git
cd sglang
# Install the python packages
pip install --upgrade pip
pip install -e "python"
运行推理
SGLang 目前同时支持 BF16 和 FP8 模型。这取决于 ${MODEL_PATH} 中模型的 dtype。
以下是在多 GPU 节点上运行 Ling-1T 的示例，其中主节点 IP 为 ${MASTER_IP}，服务器端口为 ${PORT}：
- 启动服务器：
# Node 0:
python -m sglang.launch_server --model-path $MODEL_PATH --tp-size 8 --pp-size 4 --dp-size 1 --trust-remote-code --dist-init-addr $MASTER_IP:2345 --port $PORT --nnodes 4 --node-rank 0
# Node 1:
python -m sglang.launch_server --model-path $MODEL_PATH --tp-size 8 --pp-size 4 --dp-size 1 --trust-remote-code --dist-init-addr $MASTER_IP:2345 --port $PORT --nnodes 4 --node-rank 1
# Node 2:
python -m sglang.launch_server --model-path $MODEL_PATH --tp-size 8 --pp-size 4 --dp-size 1 --trust-remote-code --dist-init-addr $MASTER_IP:2345 --port $PORT --nnodes 4 --node-rank 2
# Node 3:
python -m sglang.launch_server --model-path $MODEL_PATH --tp-size 8 --pp-size 4 --dp-size 1 --trust-remote-code --dist-init-addr $MASTER_IP:2345 --port $PORT --nnodes 4 --node-rank 3
# This is only an example. Please adjust arguments according to your actual environment.
- 客户端：
curl -s http://${MASTER_IP}:${PORT}/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "auto", "messages": [{"role": "user", "content": "What is the capital of France?"}]}'
更多用法可参见此处
微调
我们推荐你使用 Llama-Factory 来微调 Ling-1T。
许可证
本代码仓库基于 MIT 许可证 授权。
🤗 Hugging Face | 🤖 ModelScope
Introduction
Ling-2.5-1T, Inclusive Intelligence, Instant Impact.
Thinking models raise the ceiling of intelligence, while instant models expand its reach by balancing efficiency and performance—making AGI not only more powerful, but also more accessible. As the latest flagship instant model in the Ling family, Ling-2.5-1T delivers comprehensive upgrades across model architecture, token efficiency, and preference alignment, designed to bring universally accessible AI to a new level of quality.
- Ling-2.5-1T features 1T total parameters (with 63B active parameters). Its pre-training corpus has expanded from 20T to 29T tokens compared to the previous generation. Leveraging an efficient hybrid linear attention architecture and refined data strategy, the model delivers exceptionally high throughput while processing context lengths of up to 1M tokens.
- By introducing a composite reward mechanism combining "Correctness" and "Process Redundancy", Ling-2.5-1T further pushes the frontier of efficiency-performance balance in instant models. At comparable token efficiency levels, Ling-2.5-1T’s reasoning capabilities significantly outperform its predecessor, approaching the level of frontier "thinking models" that typically consume ~4x the output tokens.
- Through refined alignment strategies—such as bidirectional RL feedback and Agent-based instruction constraint verification—Ling-2.5-1T achieves substantial improvements over the previous generation in preference alignment tasks, including creative writing and instruction following.
- Trained with Agentic RL in large-scale high-fidelity interactive environments, Ling-2.5-1T is compatible with mainstream agent platforms such as Claude Code, OpenCode, and OpenClaw. It achieves leading open-source performance on the general tool-calling benchmark, BFCL-V4.
Model Downloads
| Model | Context Length | Download |
|---|---|---|
| Ling-2.5-1T | 256K -> 1M (YaRN) | 🤗 HuggingFace 🤖 ModelScope |
Note: If you are interested in previous version, please visit the past model collections in Huggingface or ModelScope.
Deployment
SGLang
Environment Preparation
We will later submit our model to SGLang official release, now we can prepare the environment following steps:
git clone -b ling_2_5 git@github.com:antgroup/sglang.git
cd sglang
# Install the python packages
pip install --upgrade pip
pip install -e "python"
Run Inference
Both BF16 and FP8 models are supported by SGLang now. It depends on the dtype of the model in ${MODEL_PATH}.
Here is the example to run Ling-1T with multiple GPU nodes, where the master node IP is ${MASTER_IP} and server port is ${PORT}:
- Start server:
# Node 0:
python -m sglang.launch_server --model-path $MODEL_PATH --tp-size 8 --pp-size 4 --dp-size 1 --trust-remote-code --dist-init-addr $MASTER_IP:2345 --port $PORT --nnodes 4 --node-rank 0
# Node 1:
python -m sglang.launch_server --model-path $MODEL_PATH --tp-size 8 --pp-size 4 --dp-size 1 --trust-remote-code --dist-init-addr $MASTER_IP:2345 --port $PORT --nnodes 4 --node-rank 1
# Node 2:
python -m sglang.launch_server --model-path $MODEL_PATH --tp-size 8 --pp-size 4 --dp-size 1 --trust-remote-code --dist-init-addr $MASTER_IP:2345 --port $PORT --nnodes 4 --node-rank 2
# Node 3:
python -m sglang.launch_server --model-path $MODEL_PATH --tp-size 8 --pp-size 4 --dp-size 1 --trust-remote-code --dist-init-addr $MASTER_IP:2345 --port $PORT --nnodes 4 --node-rank 3
# This is only an example. Please adjust arguments according to your actual environment.
- Client:
curl -s http://${MASTER_IP}:${PORT}/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "auto", "messages": [{"role": "user", "content": "What is the capital of France?"}]}'
More usage can be found here
Finetuning
We recommend you to use Llama-Factory to finetune Ling-1T.
License
This code repository is licensed under the MIT License.