Whenever an important new model is released, our editorial team usually gets to work even earlier than the news itself: we open the official website, live streams, and social media in advance, and as the results come out one after another, everyone starts writing articles with great enthusiasm, while also sorting out parameters, rankings, and demonstration cases to judge how far the technology has really gone this time.
This process has become quite sophisticated, and the large model industry seems to have developed a default aesthetic: more parameters, longer contexts, more complex reasoning, and an increasingly comprehensive list of capabilities.
But there's a classic saying in the design world:Less is MoreIt emphasizes that design should be restrained, knowing what should be retained and what can be discarded.
The same applies to models. After watching press conferences for a long time, we easily focus all our attention on the upper limit of capabilities, as if the biggest difficulty we encounter at work every day is the lack of a colleague in the office who can prove the Riemann Hypothesis.
In reality, most jobs don't concern the ultimate fate of human civilization. They are simply numerous, repetitive, have clearly defined requirements, and follow a complex workflow. In such cases, what we often need is a tool that is quick to respond, understands commands, and can be used reliably.
On July 24, Bailin Big Model, a subsidiary of Ant Group, released its latest model.Ling-3.0-flashIt was designed along this line of thinking. It is a Flash model for long-term Agent workflows, with a total of 124B parameters and 5.1B activation parameters. It emphasizes fast response, stable tool invocation, and controllable cost suitable for high-frequency use. In terms of context, Ling-3.0-flash natively supports 256K, expandable to 1M, and can maintain a high recall rate under long input.
The official positioning of it is "Wisdom prevails through secrecyCompared to the previous generation flagship thinking model Ring-2.6-1T (approximately 1T, with approximately 63B activations), Ling-3.0-flash achieved "matching or even surpassing" the flagship in most benchmark tests using only about 1/8 of the flagship's total parameters and about 1/12 of its activation parameters.
Therefore, Ling-3.0-flash is not actually telling the story of "yet another small model," but rather a shift in thinking:The way models acquire intelligence is no longer simply by pursuing large-scale accumulation, but by efficiently squeezing every inch of computing power..
Bailing called this orientation "Intelligent density"(IQ divided by memory cost) and "Smart efficiency ratio"(IQ divided by computing cost), and Ling-3.0-flash performs exceptionally well in both metrics."
From a product positioning perspective, Ling-3.0-flash does not attempt to replace those all-around inference models with a large number of parameters, but rather aims to take on more "..."implement"Work: For complex models, first search for information, formulate a plan and break down the tasks. Then, Ling-3.0-flash calls tools, modifies code, processes data according to the plan, and continuously checks and corrects based on the running results."
In this way, the parts requiring in-depth thinking can be handled by the large model, while the parts requiring repeated execution and trial and error are handled by Ling-3.0-flash. This ensures the effectiveness of the task while reducing the waiting time and costs caused by calling a huge model at every step.
This positioning aligns with an emerging industry consensus. Entering 2026, industry discussions about agents have shifted from "whether they can complete tasks autonomously" to "who should do what within an agent system." In other words,An effective agent system requires not only well-thought-out, large-scale inference models, but also relatively lightweight models that can quickly and efficiently complete basic tasks..
First-hand test
Incorporating "speed" and "stability" into real workflows
So, how does Ling-3.0-flash actually perform? Currently, Ling-3.0-flash is offering a limited-time free API call (until 23:00 Beijing time on August 3rd). Taking advantage of this window, we'll also try it out for free and test it with real-world tasks.
The learning curve was lower than expected. Our testing tool was OpenCode; after downloading and opening it, Ling-3.0-flash was already one of the default free options and could be used directly without additional configuration. If you want to integrate other tools such as Claude Code, you can apply for a free API key from the OpenRouter or Bailing official platform and configure it yourself.
Official Experience Platform:https://chat.ant-ling.com/chat
OpenRouter demo address:https://openrouter.ai/inclusionai/ling-3.0-flash:free
Our first task was to have it build a simulated solar system. The result was somewhat unexpected: Ling-3.0-flash not only correctly constructed the relative positions of the sun and the eight planets, but also actively used different colors and planetary rings to visually differentiate them. Furthermore, Ling-3.0-flash completed the task remarkably quickly, in just 3 minutes.
The second task was more aesthetically pleasing: we asked it to create a Bauhaus-style "Machine Heart Selected Portfolio" website with the theme "Less is More," without using any image materials at all.
The finished product accurately reproduces the image-free aesthetics of pure CSS/SVG, including primary colors, geometric composition, and asymmetrical grids. This can be seen as a testament to the quality of similar official demos. In this type of front-end visual task, its performance truly lives up to the description of being a "comparison."
Beyond visual presentation, we were more concerned with whether it could truly be integrated into the production process. So we integrated it into our workflow and instructed it to install "Machine Heart Data Skills," which integrates all of Machine Heart's self-produced industry data.
Soon, Ling-3.0-flash integrated these four skills into OpenCode. After configuring the API key, users can simply ask a question in the dialog box, and it will be able to query all of Machine Heart's industry data according to the methods defined in the skill.
For example, we can have it retrieve all the models and technical articles published by Ant Financial in the database, and summarize Ant Financial's technology evolution path accordingly.
Based on 24 relevant articles in the database, Ling-3.0-flash concisely outlines the main thread of Ant Financial's development over the years, from infrastructure to breakthroughs in MoE architecture, and then to the integration of full-modality and intelligent agent capabilities. The clear and logical structure serves as a valuable reference for quickly understanding this team.
This kind of work, which involves "reading a batch of structured materials, extracting key points, and organizing them into a document," is precisely the kind of information aggregation and structured output scenario that it excels at.
Of course, we didn't just stick to what it's good at. We tried to get it to complete a more complex project independently: a 3D physics-based billiards sandbox using Vite + TypeScript + Three.js and the cannon-es physics engine. This time, it didn't run completely.
This result was not unexpected; rather, it corroborated Ling-3.0-flash's own positioning: it was never designed as an all-around architect to "handle complex projects end-to-end with a single sentence," but...It is more suitable to implement the architecture and path in stages as execution nodes after people have clearly planned them out.If the modular framework of the billiards sandbox is first built by people or a larger planning model, and then filled in round by round, it is likely to be a more convenient collaborative approach.
Based on several rounds of testing, our intuitive feeling largely matches the official statement: for tasks with clear boundaries and explicit verification signals, it responds quickly, executes stably, and has low cost of integrating into the workflow; however, once the task becomes open, loose, and requires it to build the overall architecture from scratch, its shortcomings become apparent.
This is exactly what an "execution model" should look like: when used in the right place, it is very cost-effective; when used in the wrong place, it may lead to disappointment.
Behind the extreme "exploitation"
An architecture rewritten around efficiency
The official explanation for Ling-3.0-flash's capabilities is the combined effect of two lines: a systemic upgrade of the underlying architecture and targeted optimization of agents for productivity scenarios.
The former determines "how much capability can be generated from the same parameters", while the latter determines "whether these capabilities can be delivered stably in real-world tasks".
Let's look at the architecture first.Native Hybrid Linear AttentionThis is the foundation for this upgrade.
Unlike its predecessor, Ling-2.6, which performed architecture migration before continuing training, Ling-3.0 adopts a native hybrid linear attention architecture from the pre-training stage. It alternately stacks KDA linear attention layers and MLA layers in a 5:1 ratio, meaning that every 6 layers contains 5 KDA layers and 1 MLA layer. Because all components are optimized together in the same architecture from the early stages of training, the model can balance long-context processing efficiency and overall capabilities.
The KDA (Kimi Delta Attention) mentioned here refers to the linear attention mechanism proposed by Moonshot in his Kimi Linear work at the end of 2025, which has since become a common reference for the industry in exploring efficient attention.
Its idea is to introduce fine-grained diagonal gating into the state update of Delta Rule: allowing different feature channels to have their own independent retention, decay and write control, instead of using a global knob to uniformly handle the forgetting of memory.
In layman's terms, when reading long documents or large codebases, the model can more accurately maintain key information that spans paragraphs and steps, preventing new and old information from becoming a jumbled mess. Ling-3.0 replaced the previous generation's Lightning Attention with KDA precisely because of this.
Further reduction in MoE sparsity is the direct source of the "intelligence-efficiency ratio". The methodological basis for this approach comes from the previous proposal by Behringer.Efficiency LeverageandLing Scaling LawUnder the premise of achieving the same performance, a lower expert activation ratio can result in higher computing efficiency.
The lower the activation ratio, the higher the efficiency leverage (the computing power saved relative to the dense model), and this leverage increases with the computing power budget. The figure shows the estimated value at 1e22 FLOPs. Source: Ant Financial's Efficiency Leverage paper, arXiv:2507.17702
According to reports,Ling-3.0 accordingly reduces the expert activation ratio for each token from 1/32 in the previous generation to 1/64.In other words, of the 124B total parameters, only a very small portion is illuminated in each inference, driving the model's capabilities to achieve leaps with extremely low computational consumption.
The combination of native hybrid linear attention, KDA, and sparse MoE forms the architectural foundation that allows it to "challenge beyond its class" at a scale of 124B (5.1B activations).
However, there are differing opinions within the industry as to whether this "linear attention + extreme sparsity" approach can maintain quality at a larger scale. Acing AI's third-party analysis points out that the "quality alignment" of the linear attention hybrid architecture at the research scale (billions of parameters) may not be able to smoothly migrate to the forefront scale, especially in tasks such as long-context multi-hop inference, which are underestimated by standard test sets.
Of course, Ling-3.0-flash wasn't designed to be a jack-of-all-trades; it focuses its limited resources on the Agent and core inference. However, this also reminds us that the notion of "small scale comparable to a flagship product" needs to be examined within the specific task at hand, rather than being generalized.
In addition to the architectural upgrade of the model itself,Ling-3.0-flash also features specific optimizations for Agent tasks in real-world production environments, including training, inference services, and multi-agent collaboration..
According to official statements, the team will integrate Coding Agent, General Agent, and Deep Research Agent.The number of interactive training environments has expanded to over 10,000.Furthermore, it continuously improves the diversity, quality, and task difficulty of the environment. The reinforcement learning phase also introduces complete...Execution trajectory reviewandStep-by-step contribution assessmentThis provides more granular step-level training signals for long-term tasks.
These training exercises primarily enhance the model's ability to continuously plan, invoke tools, handle failure feedback, and dynamically adjust execution paths under complex constraints. Officially, compared to its predecessor, the flagship Ring-2.6-1T, Ling-3.0-flash boasts significantly improved Agent capabilities, with some metrics reaching the level of models with larger parameter scales.
MiniAppBench provides an external benchmark—this benchmark requires models to generate a fully functional app based on just one user request. The average pass rate of the 16 mainstream models participating in the evaluation was approximately 17%, while Ling-3.0-flash achieved a pass rate of 25.3%, on par with open-source state-of-the-art models with approximately twice the total number of parameters.
There are also two engineering changes that address the issue of "the longer the agent task, the more strain the service becomes":
FirstResponse speedThe more rounds of dialogue and the longer the context of an Agent task, the more likely traditional inference services will need to reprocess the context of tens of thousands of tokens if the local cache is evicted or the session is rescheduled to another node, causing the latency of the first token to rise rapidly. Ling-3.0-flash builds a cluster-level hierarchical caching system based on SGLang HiCache and Mooncake, upgrading the cache from "instance-private" to "cluster-shared", enabling existing contexts to be stored across layers, reused across nodes, and restored on demand.
Official tests show that, in long input scenarios, hitting the L3 cache can reduce TTFT by 60% to 80% or more compared to recalculation.
For scenarios such as Coding Agent and long document question answering, this can reduce context recalculation and speed up task continuation.
SecondCooperative stabilityWhen faced with complex agent tasks, single inference chains are prone to problems such as decision drift, partial misjudgments, and insufficient fault tolerance. To address these issues, Bailing has upgraded its system to include a new inference mechanism called "...".Ling Multi-AgentThe architecture is a multi-agent collaborative architecture. Different agents can perform multi-level division of labor around the task, and reduce the bias caused by a single reasoning path through cross-validation, information supplementation, collaborative error correction, and competitive "horse race".
According to the official experimental results, Ling Multi-Agent's performance on BrowseComp and BrowseComp_zh improves with the scale of collaborative computing. This indicates that in complex retrieval and deep research tasks, capability expansion depends not only on a single model but also on more efficient task delegation and result validation.
Based on these changes, the product logic of Ling-3.0-flash is self-consistent:To maximize the speed, stability, and cost of the "execution" stage, we then use engineering methods to solve the recalculation and drift problems in long-distance interactions..
Correspondingly, this trade-off also brings clear boundaries. Model resources are more concentrated on agents and core reasoning, and there is still room for improvement in terms of breadth of knowledge reserves and multilingual stability. In long dialogues or high-pressure interactions, Chinese and English may be mixed.
For users, this means some clear "unsuitability": extremely niche research tasks requiring rote memorization of vast amounts of world knowledge, loosely defined and open-ended requirements lacking "verifiable nodes," and the one-shot expectation of "doing the whole project in a single sentence"... These scenarios are better suited to larger inference models, or for humans to plan the architecture and path first, and then execute it step by step. The current version also lacks native multimodal capabilities; for tasks requiring the understanding of screenshots or audio, it needs to be used in conjunction with a multimodal model.
When "intelligence" begins to be priced by density
From Ling-3.0-flash, we can see an increasingly obvious industry shift: as agent deployment enters deeper waters, the criteria for measuring a model's value are quietly shifting from "how many parameters it has, how high it ranks on the leaderboard" to "how much it can achieve."How much reliable execution can be achieved per unit of computing power and per unit of cost?。
This is also a trade-off many companies will be making in 2026: rather than using an expensive, large-scale end-to-end model to handle everything, it's better to...Let the large model handle the thinking, and let the lightweight execution model handle the high-frequency implementation.Dynamic routing can reduce costs.
Within this coordinate system, Ling-3.0-flash provides a good solution, but whether it is applicable to more tasks and scenarios still needs further verification. Going forward, the competition between models may not only focus on their capabilities but also on their ability to use fewer resources to apply those capabilities where they are truly needed.
For today's developers, the criteria for choosing this type of model are actually quite simple: first, figure out what you want it to do, then clarify the task boundaries and acceptance criteria, and leave the rest to it to execute quickly.
Officials stated thatThe Ling-3.0-flash model weights will be officially open source after the free trial period ends.We can start to look forward to it...