Ling 3.0 Flash

Reasoning Tool calling Long context

InclusionAI Ling 3.0 Flash is a hybrid-linear sparse mixture-of-experts reasoning model with 124B total and 5.1B active parameters. This preset serves the groupwise INT4 checkpoint on two H100 GPUs through vLLM with a 256K context window, native reasoning output, and automatic tool calling.

Setup

Sign in to Baseten

uvx truss login --browser

Install the OpenAI SDK

uv pip install openai

This preset serves Ling 3.0 Flash INT4 on H100:2 with groupwise INT4 weights and tensor parallel size 2, optimized for low-latency interactive reasoning and tool-calling workloads at the model’s native 256K context.

Hardware

H100 × 2

Engine

vLLM (8bd082c274fae025… build)

Context

256K

Concurrency

Write the config

Create and move into the project directory:

mkdir ling-3.0-flash-int4-latency && cd ling-3.0-flash-int4-latency

Then create a file named config.yaml and paste the following:

config.yaml

model_name: "model:ling-3.0-flash-int4 preset:latency"

model_metadata:
  description: >-
    InclusionAI Ling 3.0 Flash INT4 is a 124B-parameter, 5.1B-active hybrid-linear
    MoE reasoning model served with vLLM on two H100 GPUs. It exposes an
    OpenAI-compatible chat endpoint with a 256K context window, native reasoning
    output, and automatic tool calling.
  repo_id: inclusionAI/Ling-3.0-flash-int4
  tags:
    - openai-compatible
  example_model_input:
    model: inclusionAI/Ling-3.0-flash-int4
    messages:
      - role: user
        content: "Explain why the sky appears blue in three concise sentences."
    stream: true
    max_tokens: 512
    temperature: 0.6
    top_p: 0.95
    top_k: 20
    chat_template_kwargs:
      enable_thinking: true

base_image:
  # Ling 3 support (model, tool parser, and reasoning parser) landed after
  # vLLM v0.27.1. Pin the first available immutable nightly commit that includes it.
  image: "vllm/vllm-openai:nightly-5a4c8d99242e9e069b604d0e9b969e77f7dd501d@sha256:8bd082c274fae025b7079498fe1da65182ba1d4c2188c0f5a68c1042c38c3695"

weights:
  - source: "hf://inclusionAI/Ling-3.0-flash-int4@959d3a48cf05d2daf5fe7bdbbd3a6bb119e359f2"
    mount_location: "/app/checkpoint/model"
    auth_secret_name: "hf_access_token"

secrets:
  hf_access_token: null

environment_variables:
  VLLM_LOGGING_LEVEL: WARNING
  VLLM_ENGINE_READY_TIMEOUT_S: "3600"

docker_server:
  start_command: >-
    vllm serve /app/checkpoint/model
    --host 0.0.0.0
    --port 8000
    --served-model-name inclusionAI/Ling-3.0-flash-int4
    --max-model-len 262144
    --trust-remote-code
    --dtype bfloat16
    --gpu-memory-utilization 0.9
    --enable-chunked-prefill
    --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE"}'
    --enable-prefix-caching
    --tensor-parallel-size 2
    --enable-auto-tool-choice
    --tool-call-parser ling3
    --reasoning-parser ling3
  readiness_endpoint: /health
  liveness_endpoint: /health
  predict_endpoint: /v1/chat/completions
  server_port: 8000

resources:
  accelerator: H100:2
  use_gpu: true

runtime:
  predict_concurrency: 32
  health_checks:
    restart_check_delay_seconds: 1800
    restart_threshold_seconds: 1200
    stop_traffic_threshold_seconds: 120

This preset deploys the groupwise INT4 checkpoint of Ling 3.0 Flash through vLLM on two H100 GPUs with tensor parallelism. It enables prefix caching, chunked prefill, and full plus piecewise CUDA graphs. The deployment exposes an OpenAI-compatible chat completions endpoint with reasoning and tool calling at the model’s 256K context window.

Flags

The start_command passes these flags to the engine. Each one controls a runtime or serving behavior:

Flag	Value	What it does
`--max-model-len`	`262144`	Maximum context length (tokens) the server accepts per request.
`--trust-remote-code`	(no value)	Execute model-specific Python from the checkpoint (required for many Qwen, Phi, and custom architectures).
`--dtype`	`bfloat16`	Weight precision loaded at runtime. bfloat16: BF16 weights, no quantization.
`--gpu-memory-utilization`	`0.9`	Fraction of GPU memory vLLM may use for weights and KV cache.
`--enable-chunked-prefill`	(no value)	Process long prompts in chunks so decode requests keep running.
`--compilation-config`	`{"cudagraph_mode":"FULL_AND_PIECEWISE"}`	vLLM compilation passes (op fusion, dead-code elimination).
`--enable-prefix-caching`	(no value)	Reuse KV cache across requests that share a prefix.
`--tensor-parallel-size`	`2`	Number of GPUs to shard the model across.
`--enable-auto-tool-choice`	(no value)	Let the model choose when to call tools without requiring `tool_choice: "required"`.
`--tool-call-parser`	`ling3`	Server-side parser that emits structured `tool_calls` on the response. ling3: Ling 3.0 Flash tool format.
`--reasoning-parser`	`ling3`	Server-side parser that separates reasoning output into `reasoning_content`. ling3: Ling 3.0 Flash thinking format.

Deploy

Push the config to Baseten:

uvx truss push

You should see output similar to:

✨ Model ling-3.0-flash-int4-latency was successfully pushed ✨

   Model ID:      abc1d2ef
   Deployment ID: xyz123
   Endpoint:      model-abc1d2ef.api.baseten.co
   Logs:          https://app.baseten.co/models/abc1d2ef/logs/xyz123

truss push prints your model ID (abc1d2ef in the example). The examples below use it wherever you see {model_id}, and read your API key from the BASETEN_API_KEY environment variable.

Call the model

Your deployment serves an OpenAI-compatible API. Now call your deployment to run inference:

Python
cURL

main.py

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["BASETEN_API_KEY"],
    base_url="https://model-{model_id}.api.baseten.co/environments/production/sync/v1",
)

response = client.chat.completions.create(
    model="inclusionAI/Ling-3.0-flash-int4",
    messages=[
        {"role": "user", "content": "What is machine learning?"}
    ],
)

print(response.choices[0].message.content)

curl -s https://model-{model_id}.api.baseten.co/environments/production/sync/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $BASETEN_API_KEY" \
  -d '{
    "model": "inclusionAI/Ling-3.0-flash-int4",
    "messages": [
      {"role": "user", "content": "What is machine learning?"}
    ]
  }'

The server parses the model’s chain of thought into a separate reasoning_content field on the response. Read it alongside the final answer:

response = client.chat.completions.create(
    model="inclusionAI/Ling-3.0-flash-int4",
    messages=[
        {"role": "user", "content": "How many r's in strawberry?"}
    ],
)
print(response.choices[0].message.reasoning_content)  # chain of thought
print(response.choices[0].message.content)            # final answer

To let the model call tools, pass a tools array. The server returns structured tool_calls on the response:

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "parameters": {
            "type": "object",
            "properties": {"location": {"type": "string"}},
            "required": ["location"],
        },
    },
}]

response = client.chat.completions.create(
    model="inclusionAI/Ling-3.0-flash-int4",
    messages=[
        {"role": "user", "content": "What's the weather in Paris?"}
    ],
    tools=tools,
)
print(response.choices[0].message.tool_calls)

Examples

Models

Engines

Custom Docker servers

Custom Python models

Chains

Setup

Hardware

Engine

Context

Concurrency

Write the config

Flags

Deploy

Call the model

Next steps

Call your model

Autoscaling

​Setup

Hardware

Engine

Context

Concurrency

​Write the config

​Flags

​Deploy

​Call the model

​Next steps

Call your model

Autoscaling

Setup

Write the config

Flags

Deploy

Call the model

Next steps