Faster serving and inference, more flexible deployment, bringing intelligence into real business
Open intelligence for the real-world.
Bailing's new-generation native hybrid reasoning model Ling-3.0-flash is now officially open-sourced.
We are simultaneously providing the base version along with FP8, FP4, INT4 and other versions, so developers can choose the deployment approach that suits them based on model precision, hardware conditions, inference cost and data security requirements.
Ling-3.0-flash adopts a MoE architecture with 124B total parameters and 5.1B activated parameters. It retains the knowledge and capability capacity of a hundred-billion-parameter model, but activates only about 5.1B parameters when generating each token, thereby reducing the computational overhead of a single inference.
This makes Ling-3.0-flash not just a model that "runs fast," but one that can genuinely enter business systems of different scales, offering three different paths to deployment:
- API access, letting Agent applications go live quickly: For developers who want fast integration without building their own inference services, Ling-3.0-flash can be called directly via cloud API;
- Single-machine private deployment, keeping data local: For enterprises and teams whose data cannot leave their domain, the MXFP4 and INT4 versions can complete end-to-end inference on a single NVIDIA DGX Spark;
- High-performance deployment, unleashing extreme single-request speed: For high-performance services sensitive to single-request latency, under the specified GPU test configuration, the average output rate exceeds 1100 tokens/s.
Developers can either call the model directly or freely validate, deploy and customize model capabilities, integrating Ling-3.0-flash into their own business systems, development toolchains and Agent workflows.
From cloud APIs to a single DGX Spark, AMD 395 and domestic inference hardware, and on to high-performance inference clusters, Ling-3.0-flash aims to give more teams deployable, sustainable and controllable intelligence.
01 Direct API access: getting high-frequency Agents online faster
For developers and product teams who want to quickly validate products and launch Agent applications, the API is the lowest-barrier integration method. Without first building and maintaining inference infrastructure, Ling-3.0-flash can be integrated into workflows such as code development, search and research, and tool calling.
On the Artificial Analysis leaderboard, Ling-3.0-flash achieves an output speed of 353 tokens/s. A higher sustained output speed can directly shorten waiting time in code generation, long-text processing and Agent execution.
This value is especially evident in multi-turn Agent workflows.
Completing an Agent task typically requires going through multiple rounds of "understand the task—generate a plan—call tools—read results—correct actions." When each round of generation is faster and fewer tokens are used to complete the task, cumulative waiting time and calling costs both decrease.
Therefore, Ling-3.0-flash is especially suited for: Coding Agents and code generation tools; search, research and information integration Agents; API, MCP and terminal tool calling; web, game and lightweight application development; office documents and data processing; and execution nodes in multi-agent collaboration.
On the AA Intelligence Index leaderboard, Ling-3.0-flash has a weighted average calling cost of about $0.04 per task and a weighted average decoding time of about 1.4 minutes, placing it in the advantageous region of both "intelligence level—task cost" and "intelligence level—task duration."
This shows that the significance of 353 tokens/s is not just that "words come out faster," but that a complete task can be advanced at lower cost and in less time.
02 Single-machine local deployment: keeping sensitive data local
From open weights to single-machine deployment: keeping data and the technology path in your own environment
Recently, when discussing the significance of open-weight models for the AI ecosystem, Sebastian Raschka pointed out that open-weight not only allows external researchers to validate models and test vendors' technical claims, but also lets developers and enterprises run AI on their own hardware, reducing the need for personal data, business information and intellectual property to enter external model service pipelines. In the same tweet, he also specifically mentioned that the technical community is waiting for Kimi K3 and Ling 3.0 weights to go live on model platforms.
The community is waiting for a set of open models that are truly obtainable, runnable and verifiable, which is also an important reason why Ling-3.0-flash is open-sourcing multiple precision versions this time and continuously improving local deployment capabilities. For teams more focused on data security and deployment autonomy, local operation offers a more controllable option. Through the MXFP4 and INT4 quantized versions, Ling-3.0-flash has already completed end-to-end inference validation on a single NVIDIA DGX Spark equipped with 128GB LPDDR5x memory.
In local tests, the Ling-3.0-flash MXFP4 version achieves a generation speed of about 80 tokens/s; in long-text input scenarios, Prefill speed can reach about 2500–3500 tokens/s. Under the test conditions, a single Spark can support about 3–4 users.
🎬 Video demo (seethe original WeChat article ）
This means that capabilities of hundred-billion-parameter models that previously typically required multi-GPU servers to run can now be validated and used on a single desktop-class AI device.
Single-machine deployment allows users to build a local intelligence system that is runnable, verifiable and sustainably usable without relying on large data centers. Its value is mainly reflected in three aspects:
- Data stays local, making sensitive business more controllable. The model, business data and inference process can run in a local closed loop, reducing the risk of sensitive information being transmitted externally and handled by third parties. Suitable for scenarios such as enterprise internal documents and code, medical research materials, financial data and legal materials.
- Reduced external dependence, making long-term tasks more stable. Once the model, inference framework and business tools are all deployed locally, the system can run on an intranet or in an offline environment, reducing the impact of network fluctuations, API quotas and changes in third-party services. Suitable for 24/7 background Agents, batch document analysis, automatic generation and fixed-process automation.
- Integration with internal systems, so even small teams can truly put it to use. Local models can connect to an enterprise's own permission systems, knowledge bases, business systems and operational processes, rather than serving only as an external chat tool. Under measured conditions, a single DGX Spark can support about 3–4 users, suitable for department-level knowledge assistants, R&D team Coding Agents, research material analysis and office automation.
Medical research data processing on a single Spark: a local closed loop for sensitive data
We deployed Ling-3.0-flash on a single NVIDIA DGX Spark and built a local Agentic Pipeline for medical research data processing. Both the model and the data run in a local environment, reducing the risk of patient information entering external pipelines, and can assist with:
-
Sensitive information processing: de-identifying identity information and assigning dynamic IDs to research samples, reducing the risk of linking back to real identities during subsequent processing.
-
Structured medical information: extract vital signs, blood biochemistry indicators, ultrasound reports and other information from raw materials, and organize scattered data into structured results that can be used for research analysis.
-
Research workflow assistance: complete preliminary sample grouping according to research rules; for samples that are abnormal, complex, or lack sufficient evidence, uniformly mark them as "Unclassified" and hand them over to professionals for further review.
🎬 Video demonstration (seethe original WeChat article ）
This case demonstrates the practical value of single-machine deployment of Ling-3.0-flash: while reducing the risk of sensitive medical data being transmitted externally, it transforms repetitive information extraction, organization, and preliminary classification into a continuously operating local workflow, allowing researchers to spend more time on work that truly requires professional experience and responsible judgment.
The model is used to assist medical research and data processing, and does not replace professional judgment or clinical diagnosis. Practical application still needs to meet relevant regulations, ethical review, and organizational requirements for permission management, log auditing, and data governance.
03 Dedicated high-performance deployment: making real-time Agents wait less
Ling-3.0-flash lowers the entry threshold for private deployment through single-machine deployment; for real-time businesses that are more sensitive to single-request latency, Ling-3.0-flash further compresses generation wait time through deep collaboration between high-performance GPUs and the inference stack, providing a lower-latency operating experience for real-time Agents, multi-turn tool calls, and continuous task execution.
In a test environment with TP4, Batch Size 1, input length 8192, and output length 1024, through inference system optimization and DSpark speculative decoding, Ling-3.0-flash achieved an average output rate exceeding 1100 tokens/s, with average TPOT as low as about 0.8ms.
This capability is mainly aimed not at ordinary batch services, but at real-time interactive scenarios that are extremely sensitive to single-response speed.
For example: real-time code completion and code modification in IDEs; Agents that need continuous observation, decision-making, and operation; browser and UI automation; multi-turn terminal operations and fault repair; workflows that frequently call search, databases, and business APIs; professional research assistants that are sensitive to response wait time.
For such applications, the model often does not generate an answer once, but performs multiple rounds of reasoning and tool calls within a very short time. The lower the single-request latency, the more observations and corrections the Agent can complete per unit time, and the more coherent the entire task flow becomes.
To reduce waiting and GPU bubbles in Batch-1 scenarios, we mainly carried out two types of optimization.
- The first is inference engineering optimization. Through CPU/GPU overlap, operator-level overlap, KDA operator optimization, and operator fusion, we reduce CPU synchronization and waiting between operators, enabling the GPU to maintain higher effective utilization in single-concurrency scenarios.
- The second is speculative decoding. The DSpark Draft Model trained for Ling-3.0-flash can generate a group of candidate tokens at once, which are then centrally verified by the target model. For general scenarios, a more robust Block Size can be selected; for vertical businesses with relatively stable patterns, business data can also be combined to improve the Draft Model's prediction accuracy and further optimize speed.
It should be noted that 1100+ tokens/s is the average output rate for Batch-1 under a specific software and hardware configuration. It mainly reflects the low-latency potential of the model architecture and inference system, and is not equivalent to the total throughput of a high-concurrency server. Network time, time to first token, and external tool execution time are also not included in this figure.
For businesses, what this set of data really shows is: Ling-3.0-flash can become a high-speed execution engine in real-time Agent workflows.
04 Three deployment methods: how should you choose?
If your goal is to quickly integrate the model, validate a product, or handle elastic usage, we recommend prioritizing the API.
If data cannot leave the domain, the number of users is relatively limited, or you need to run continuously on an internal network, you can choose the MXFP4 or INT4 version and complete local deployment on devices such as DGX Spark.
If your business requires extremely low single-request latency, such as real-time coding assistants, high-frequency Agents, and continuous tool calls, you can use high-performance hardware for dedicated deployment and combine it with inference optimization capabilities such as DSpark.
The same Ling-3.0-flash can be adopted in different ways according to the business stage:
- use the API during the product validation stage;
- move sensitive data and internal workflows to local private deployment;
- build high-performance dedicated services after business growth.
Developers do not need to bear the full infrastructure cost from the start, nor do they need to be locked into a single calling method.
05 Open source: letting the model truly enter business systems
During the open experience period of the Ling-3.0-flash API, community developers conducted extensive exploration in tasks such as code generation, web and game development, Web3 research, multi-agent collaboration, API/MCP tool calls, and UI automation.
The most frequently appearing evaluation in feedback is "efficient and reliable executor."
Ling-3.0-flash can decompose and advance tasks according to context, handle tool call errors and retries, maintain task state in long-running processes, and reduce loop stalls and repeated execution.
Compared with taking charge of an entire project in one go starting from vague requirements, Ling-3.0-flash is better suited as an execution node in Agent workflows: when goals, boundaries, and processes are clear enough, it can balance output quality, token efficiency, and execution speed, providing a stable and lightweight connection between "planning" and "execution."
During model ecosystem adaptation, we also conducted collaborative testing and adaptation with Agent and development tool partners such as OpenCode, Kilo Code, and nanobot, inference service providers such as Novita AI, Fireworks AI, and DeepInfra, and open-source inference frameworks such as SGLang.
From API to inference frameworks, from development tools to local devices, Ling-3.0-flash is forming a more complete application and deployment ecosystem.
Bailing Real-World Intelligence Co-Creation Program — with 10 billion tokens, bringing open intelligence into real business
The value of open intelligence is not only reflected in model leaderboards, but should be validated in real needs, real processes, and real users.
The "Bailing Real-World Intelligence Co-Creation Program" is now open for a limited time to developers, AI entrepreneurs, enterprise practitioners, and open-source contributors. We will recruit 100 co-creators and provide each with 100 million tokens free of charge, totaling 10 billion tokens, to support everyone in applying Ling-3.0-flash to real business and actual workflows.
Whether you are building Agents, developing AI products, maintaining open-source projects, or hoping to use large models to transform specific processes such as code development, knowledge management, office collaboration, and research analysis, you are welcome to submit an application. What we look forward to seeing is not just an idea, but a clear problem, a feasible plan, and results that can be used and validated by real users.
Each application will be carefully evaluated. Limited resources will be prioritized for projects with genuine needs, clear application scenarios, and sustained execution plans. We also hope to work with selected co-creators to explore how open models can move from technical capability to business value, turning models into runnable products, reusable workflows, and verifiable practical outcomes.
Sign up now and join us in bringing open intelligence into the real world:
Go to the Ling official website to apply
https://ant-ling.app.weavefox.cn/
Limited-time discount, open experience continues
With the end of the two-week limited-time free trial, to further reduce developers' ongoing usage costs,Ling Studio is offering a limited-time 75% off discount on Ling-3.0-flash. Developers can also continue to access the model conveniently through the Ling official platform and third-party platforms such as OpenRouter, and further explore its capabilities in scenarios such as complex task execution, tool calling, code generation, and long-context understanding.
Note: From 00:00 on August 7, 2026 to 24:00 on August 31, Ling-3.0-flash enjoys a limited-time 75% off discounton Ling Studio; the original price will be restored from 00:00 on September 1.
Open-source weights and local deployment
We welcome everyone to try it, give feedback, and exchange ideas
🤗 Hugging Face
https://huggingface.co/inclusionAI/Ling-3.0-flash
https://huggingface.co/inclusionAI/Ling-3.0-flash-fp8
https://huggingface.co/inclusionAI/Ling-3.0-flash-fp4
https://huggingface.co/inclusionAI/Ling-3.0-flash-int4
🤖 ModelScope
https://modelscope.cn/models/inclusionAI/Ling-3.0-flash
https://modelscope.cn/models/inclusionAI/Ling-3.0-flash-fp8
https://modelscope.cn/models/inclusionAI/Ling-3.0-flash-fp4
https://modelscope.cn/models/inclusionAI/Ling-3.0-flash-int4
Model release technical sharingPrevious post: Ling-3.0-tiny is officially open-sourced: bringing 1.3B activated parameters into real tasksNext post: Ling-3.0-flash: Wisdom wins through density