Ling-3.1-flash-1019

Ling-3.1-flash-1019

Handles scientific and agentic tasks with 560B total and 25B active parameters and a 1M-token context, tuned further from Ling-3.1-flash. Serving examples use SGLang and vLLM on multi-GPU servers, under MIT.

Made by inclusionAI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Ling-3.1-flash-1019 builds on Ling-3.1-flash, adds more post-training, and comes from inclusionAI. It has 560B parameters in total, uses 25B for each token, and reads a context of 1M tokens. Extra post-training raised its scores on scientific and agentic tests, which the card reports as better multi-step work and tool-driven workflows.

Who it suits: It suits teams with multi-GPU servers who want an open model for agents, scientific questions and office automation.

Industry | Who uses it | How they use it | Example
Software and AI teams | Teams that serve large models | Run the model behind SGLang or vLLM | Start a server and send chat requests with thinking on
Office automation | Teams that build agent workflows | Let the model plan and call tools over several steps | Run a task from the AutomationBench-AA style of work
Science research | Researchers who test models on science questions | Compare it with its predecessor | Look at the CritPt and SciCode gains the card lists

If you do not have a multi-GPU server, you can skip it, since even the 4-bit weights are about 286 GB.

Where to use it:

• SGLang, which the card recommends, through a Docker image
• vLLM, with its own environment and server settings
• The ModelScope and OpenRouter pages that the card links to
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Agent workflows | Developers | Multi-step task handling with tool calls
Scientific questions | Researchers | Answers on tests such as CritPt and SciCode
Tool-driven office tasks | Automation teams | Results from tasks that need tools and reasoning

How to set it up: What you need first:

• Docker with GPU access for the SGLang route, or uv and a Python environment for vLLM
• A multi-GPU server, such as 8 x H200 for the card's example
• Enough disk for about 572 GB of FP8 weights

1. Pull the SGLang image that supports this model family.
2. Start the server with the model path and the parallel settings for your GPUs.
3. Or prepare a vLLM environment and start its server instead.

docker pull lmsysorg/sglang:dev-ling-3.1-flash

docker run --gpus all --network host --ipc=host --rm \
  lmsysorg/sglang:dev-ling-3.1-flash \
  python3 -m sglang.launch_server \
    --model-path inclusionAI/Ling-3.1-flash \
    --tp 8 --ep 8 \
    --speculative-algorithm NEXTN \
    --host 0.0.0.0 --port 30000

For vLLM, create a uv environment first.

pip install uv

uv venv ~/my_ling_env

source ~/my_ling_env/bin/activate

Next, clone the branch named vllm from the vLLM project on GitHub, then install it from inside that folder.

cd vllm

VLLM_USE_PRECOMPILED=1 uv pip install --editable . --torch-backend=auto

To fetch this repository's files yourself, use the Hugging Face tools.

pip install -U huggingface_hub
hf download inclusionAI/Ling-3.1-flash-1019

How to use it: Start a server, then send chat requests to it. The card's SGLang example names the Ling-3.1-flash weights, so check which weights you load. For the best results, set temperature 1.0, top_p 0.95 and top_k 20, and turn thinking on.

vllm serve "$MODEL_PATH" \
    --port "$PORT" \
    --trust-remote-code \
    --served-model-name auto \
    --tensor-parallel-size 8 \
    --gpu-memory-utilization 0.85 \
    --enable-prefix-caching \
    --enable-expert-parallel \
    --mamba-cache-mode align \
    --enable-auto-tool-choice \
    --tool-call-parser ling3 \
    --reasoning-parser ling3 \
    --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

Tool | What it does
SGLang | Serves the model with NEXTN speculative decoding and Hopper decode optimizations
vLLM | Serves it with tool call and reasoning parsers named ling3, plus MTP speculative decoding

Example: You start the vLLM server on 8 GPUs and ask a short factual question with thinking on. The server answers in the OpenAI-style chat format, with the reasoning split out by the ling3 parser.

For newcomers and experts: This model needs a GPU cluster, so a first try is easier through the hosted pages the card links to. If you run it yourself, start with one short question and the recommended sampling settings.

For experts, the design is a hybrid-linear mixture of experts with 63 KDA layers and 9 gated MLA layers in a 7 to 1 ratio, plus 4 dense layers. It has 512 routed experts, 1 shared expert and 8 activated experts per token, with 64 attention heads and a hidden size of 4096. The trade is memory: only 25B parameters work per token, but all 560B must be loaded.

Prepare SGLang or a vLLM environment
Start the server with parallel settings
Send a chat request with thinking on
Get the reply with reasoning separated

Good to know:

Title: Artificial Analysis Intelligence Index v4.3.2
Ling-3.1-flash: 41
Ling-3.1-flash-1019: 44

Title: Reported gains over Ling-3.1-flash
Benchmark | Ling-3.1-flash | Ling-3.1-flash-1019
AA-Briefcase | 1397 | 1524
AutomationBench-AA | 62% | 68%
CritPt | 18% | 26%
SciCode | 54% | 57%

• Index v5: the newer Artificial Analysis index, version 5, gives it 30 according to the card.
• Hardware: the example server uses 8 x H200 with TP8 and EP8.
• Size: the files weigh about 572 GB in FP8 form, and about 286 GB at 4-bit, which is an estimate.
• Scores: the authors report all benchmark numbers themselves.
• License: mit
• Cost to run: no API key or paid service named. The GPU hardware is the cost.
• Data it reaches: it pulls a Docker image, a vLLM branch and the model files, and the example server listens on every network interface on port 30000.
• Not for: small machines, since the example needs 8 x H200.

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 572B
• Weights: about 572 GB in F8_E4M3; about 286 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,304
• Likes: 99
• License: mit
• Quantized versions: 14
• Tags: safetensors, bailing_hybrid, text-generation, conversational, fp8

Links

• Model: https://huggingface.co/inclusionAI/Ling-3.1-flash-1019
• Files: https://huggingface.co/inclusionAI/Ling-3.1-flash-1019/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download inclusionAI/Ling-3.1-flash-1019

Then follow the usage in its model card. Nothing in this kit runs by itself.