Today, we are officially open-sourcing Ling-3.0-tiny.
Ling-3.0-tiny is a native hybrid reasoning model with 7.9B total parameters and only 1.3B activated parameters at inference time. This open-source release simultaneously provides BF16, FP8, and INT4 versions, allowing developers to flexibly choose based on hardware environment, capability requirements, and deployment cost.
We are releasing Ling-3.0-tiny not to join a simple parameter-scale race.
What we truly care about is whether, under more limited computational resources, a model's capability, efficiency, and deployability can all hold true at the same time:
- With only 1.3B parameters activated per token, can it still possess comprehensive capabilities covering real-world tasks?
- Can a lightweight model balance complex task reasoning, Agent execution, and fast response?
- Can a more efficient architecture truly translate into lower inference costs and a lower deployment barrier?
- Can small and medium-sized teams and individual developers connect the model to their local data, tools, and business processes on their own devices?
These questions together define Ling-3.0-tiny: carrying more complete task capabilities with fewer activated parameters, and letting open intelligence truly enter real environments through more flexible precision and device choices.
01 1.3B Activated Parameters: Entering the Capability Competition Range of Mainstream Lightweight Models
For a Tiny model, fewer parameters are only the starting point. What truly matters is whether, after greatly reducing inference computation, the model can still retain the comprehensive capabilities needed to solve real-world tasks.
In the Artificial Analysis Intelligence Index evaluation system, Ling-3.0-tiny scores 25. This index comprehensively examines a model's performance across nine directions—real-world tasks, tool use, code, scientific reasoning, knowledge reliability, and long context—reflecting the model's cross-task comprehensive capability.
As shown in the figure below, Ling-3.0-tiny's comprehensive index: it only trails Gemma-4-26B-A4B's 26 points, approaching larger MoE models with a 4B activation scale; it is higher than gpt-oss-120B (high), Qwen3.5-9B, Gemma-4-12B, and Gemma-4-E4B.
Ling-3.0-tiny has 7.9B total parameters, but activates only 1.3B parameters per token. With a lower activation scale, it not only enters the comprehensive capability range of mainstream 4B–12B-class models, but also shows the potential to compete with larger-scale models.
This does not mean a single score can summarize all of a model's capabilities, nor does it mean Tiny models can replace all large models. What it proves is that Ling-3.0-tiny is not a model shrunk merely to "fit," but one that, while maintaining lightweight computational characteristics, has the capability foundation to enter code assistance, knowledge workflows, tool calling, and local Agent applications.
02 From Agent Execution to Efficient Response: More Balanced Capabilities
Ling-3.0-tiny's most prominent capability direction is real-world tasks and Agent execution. Its Artificial Analysis Agentic Index reaches 16, even exceeding Gemma 4 31B's capability score; among these, GDPval-AA v2 reaches 772 Elo and τ³-Banking reaches 20.80, demonstrating the model's capabilities in task understanding, tool calling, and process advancement.
Within the scope of models compared in this release, Ling-3.0-tiny achieves leading results on IMO-AnswerBench and the non-hallucination rate metric, and shows relatively balanced capabilities in mathematics, scientific reasoning, code Agent, instruction following, and long-context tasks.
While maintaining a 25-point comprehensive intelligence index, Ling-3.0-tiny outputs at over 160 tokens/s, with an end-to-end time of about 18 seconds for 500 output tokens, including model thinking time. A smaller activation scale thus further translates into lower task waiting and faster Agent execution.
03 Behind the Extremely Low Barrier: From Architectural Efficiency to Task Efficiency
Ling-3.0-tiny continues the native hybrid linear attention approach of the Ling-3.0 series and is further optimized for lightweight and low-barrier scenarios, keeping total parameters at 7.9B and activated parameters per token at only 1.3B.
Ling-3.0-tiny's architecture design is not aimed at pursuing a single technical metric, but at putting efficiency into real-world tasks.
- The 3:1 KDA–MLA architecture (every 4 layers contain 3 KDA layers and 1 MLA layer) enables the model to process long context more efficiently;
- 128-expert sparse MoE, with each token activating 8 routed experts and 1 shared expert, enables the model to carry more complete capabilities with 1.3B activated parameters;
- Native hybrid reasoning enables fast responses for routine tasks and multi-step thinking for complex tasks within the same model;
- The Multi-Token Prediction training objective provides the foundation for more efficient token prediction and subsequent inference acceleration.
Together, these designs ultimately point to three goals:
- Reduce the computational resources required for inference;
- Lower the barrier to local deployment and secondary development;
- Enable lightweight models to enter real Agent workflows.
04 From DGX Spark to Mac: Letting Lightweight Intelligence Truly Run Locally
To lower the barrier to model deployment and application, Ling-3.0-tiny simultaneously open-sources three versions—BF16, FP8, and INT4—allowing developers to choose a more suitable deployment method based on hardware conditions, performance requirements, and business costs. Among them:
- BF16 is suitable for research evaluation and production applications pursuing full precision;
- FP8 strikes a balance among model quality, running speed, and resource usage;
- INT4 further lowers the deployment barrier, allowing more resource-constrained local environments to use the model's capabilities.
The value of open source lies not only in providing more model versions, but in letting models truly enter developers' own devices, data environments, and business processes.
DGX Spark: One Device to Support Small-Scale Stable Service
Ling-3.0-tiny can complete FP8 inference on a single NVIDIA DGX Spark, with a model weight file of about 7.85GB. In a 2K input test, the steady-state decoding speed for a single request is about 100–105 tokens/s; with two concurrent requests, the aggregate decoding throughput is about 161 tokens/s.
This means developers and small and medium-sized teams do not need to build large-scale computing clusters; with just one DGX Spark, they can build an independently running local model service and provide stable inference capabilities for a small number of users.
From enterprise internal knowledge assistants and R&D team code assistants to task agents for specific business processes, teams can more quickly complete model application validation and service launch while controlling data boundaries and deployment costs.
Live Demo: Mobile Device Automation Based on DGX Spark
We deployed Ling-3.0-tiny on Spark to drive mobile devices, demonstrating the model's ability to understand tasks, call tools, and complete automated operations and validation. For common large-batch trial runs and regression validation scenarios in development testing, local inference can balance running cost, data controllability, and execution reliability, providing a lightweight, reusable practice for mobile automation testing.
🎬 Video demo (seeOriginal WeChat article ）
MacBook: Bringing local intelligence into everyday work
Ling-3.0-tiny has now completed local runtime adaptation for MacBook (Apple Silicon), and both the BF16 and FP8 versions can run on MacBook.
On a MacBook, Ling-3.0-tiny can serve as a portable personal productivity assistant for code assistance, document processing, local knowledge Q&A, and workflow automation. Both the model and data can remain local, which reduces reliance on cloud services and offers a more controllable option for handling privacy-sensitive information.
Among these, the FP8 version further reduces the resources required for local operation and improves continuous generation speed by about 30%, enabling developers to achieve smoother multi-turn interaction and Agent experiences at lower resource cost.
Live Demo: Recreating Infinite Wiki on a MacBook Pro (36GB memory)
To further validate Ling-3.0-tiny's performance in high-frequency local interaction scenarios, we recreated the core experience of Infinite Wiki on a MacBook Pro: while reading, clicking any word prompts the model to generate a concise explanation card based on the current context; words in the card can also be clicked further, allowing knowledge exploration to continue extending downward.
🎬 Video demo (seeOriginal WeChat article ）
The entire demo runs inference locally on macOS with Ling-3.0-tiny, without calling cloud model services. In testing, the first-token response was under 100 milliseconds, and continuous clicking and multi-level drill-down remained coherent and smooth; reading paths, query records, and generated content can also remain on the device, avoiding per-call cloud API fees.
Infinite Wiki demonstrates the practical value of Ling-3.0-tiny in high-frequency on-device interaction scenarios: it can not only run on a Mac, but also deliver a response experience close to that of a native application, becoming a local knowledge engine that is always available.
Mac mini: Keeping local intelligence continuously online
The same Apple Silicon deployment path can also extend to Mac mini.
Unlike the MacBook, which focuses on mobile work, the Mac mini is better suited as a low-power, always-on local intelligence node, continuously providing knowledge base retrieval, document analysis, automated tasks, and local model API services for individuals or small teams.
It can be placed in an office, studio, or home environment, turning Ling-3.0-tiny from a model that needs to be started temporarily into a continuously online local intelligence capability that can be invoked at any time.
Live Demo: A local reading and writing assistant based on Mac mini
We deployed Ling-3.0-tiny on a Mac mini to demonstrate everyday assistance capabilities such as word-selection translation, grammar correction, and text rewriting. The model can provide a low-latency experience of answering as soon as text is selected without uploading data to the cloud, supporting offline use while protecting content privacy and offering individuals and enterprises a lightweight, low-cost local reading and writing solution.
🎬 Video demo (seeOriginal WeChat article ）
From professional-grade local AI workstations to personal computers, different precision versions and device forms together expand the application space of Ling-3.0-tiny. Developers can integrate the model into local knowledge bases, code assistants, UI automation, enterprise task agents, and more real workflows according to hardware conditions, data boundaries, and business needs.
05 Next steps: making lightweight models more reliable and easier to deploy
Next, we will continue to advance Ling-3.0-tiny in the following directions:
- Further enhance long-tail knowledge coverage and factual accuracy;
- Improve the stability of complex tool calls and long-horizon Agent tasks;
- Continue optimizing code, scientific reasoning, and long-context capabilities;
- Improve the balance among reasoning capability, response time, and task cost;
- Improve hardware adaptation, inference framework support, and deployment documentation for the FP8 and INT4 versions;
- Build a lightweight model ecosystem together with hardware vendors, universities, and the developer community.
06 Open intelligence born for the real world
The open-sourcing of Ling-3.0-tiny is not just a public release of model weights, but also our answer to how intelligence can enter the real world.
We hope that the intelligence represented by Ling-3.0-tiny has three characteristics:
- Open(Open and usable): enabling more teams to use and deploy the model and continue innovating on that basis.
- Efficient(Efficient and sustainable): enabling high-quality reasoning, tool calls, and Agent tasks to run continuously with fewer resources and lower cost.
- Professional(Professional and trustworthy): respecting real business processes, data boundaries, and professional knowledge, while always preserving human judgment and responsibility.
Fewer activated parameters do not mean that only simpler tasks can be completed. On the contrary, they give reasoning, tool calling, and Agent capabilities the opportunity to enter more devices, serve more teams, and play a role in more real-world scenarios.
Open intelligence for the real world.
Make intelligence open, make efficiency accessible, and let more teams have deployable, sustainable, and controllable intelligence.
Open-source weights and local deployment
We welcome everyone to try it, provide feedback, and exchange ideas
🤗 Hugging Face
https://huggingface.co/inclusionAI/Ling-3.0-tiny
https://huggingface.co/inclusionAI/Ling-3.0-tiny-fp8
https://huggingface.co/inclusionAI/Ling-3.0-tiny-int4
🤖 ModelScope https://modelscope.cn/models/inclusionAI/Ling-3.0-tiny
https://modelscope.cn/models/inclusionAI/Ling-3.0-tiny-fp8
https://modelscope.cn/models/inclusionAI/Ling-3.0-tiny-int4
Model release technical sharingPrevious article: From running to training: Ling-3.0-tiny × ASystem AReno, closing the single-machine Agentic RL loopNext article: Ling-3.0-flash open-sourced: open intelligence born for the real world