This Singles’ Day, Baoling’s new model gets the job done hassle‑free.

This Double Eleven, Baoling’s new model gets the job done without a headache

To those working in e‑commerce, are you ok? My friend is already falling apart, bombarding me with messages every day:

Any good AI work‑tool recommendations? Urgent, urgent, urgent—preferably one that works well and can last a long time.

You know, I actually discovered a treasure recently.

AntBaiLing’s newly open‑sourced Ling-3.1-flash, I rely on it for my weekly work reports and topic summaries.

u1s1, I’ve tried Claude Code, Codex, and the like, and in the end I found that domestic models win my heart.

Especially Baoling, which stays low‑key, doesn’t chase parameters or rankings, but when it’s time to get work done, it’s incredibly reliable.

According to the official description, this model focuses on one thing:Getting the real‑world workflow tasks done.

Take the e‑commerce report at the beginning as an example, the boss says “Put together the details of this big promotion”, and it falls on the employee to gather information, verify data, track progress, and then create an Excel file and a presentation PPT…

Behind a single business often requires coordinated effort across multiple steps, and the hassle is that the data is scattered across different files and tools.

△Image source: Bailing official WeChat

Ling-3.1-flash is a professional model with a high intelligence‑efficiency ratio derived from these real needs.High intelligence‑efficiency ratio professional model.

In plain terms, it aims to prove that delivering a beautiful answer is secondary; getting the task done is what truly matters.

Built for Real‑World Tasks

To handle fragmented work like large‑scale promotion analysis, a model must continuously advance across multiple processes. The design of the Ling-3.1-flash model can be understood along this task chain.

The first problem the model needs to solve isto break down the delivery goal clearly.

The user may only give a vague instruction, but what actually needs to be delivered is a complete set of reporting materials. The model must find the relevant information itself, determine what’s missing and what’s needed, and then continue forward based on the intermediate results.

Ling-3.1-flashThis is exactly how it works during operation.

It links requirement understanding, task decomposition, information integration, tool invocation, and result verification, enabling work across web pages, documents, spreadsheets, and code environments.

It can create Excel and PPT files, and also generate and test code.

Moreover, this set of capabilities already has a fairly concreteimplementation form.

Ling-3.1-flash has been integrated into Ant Group’s internal desktop collaboration Agent client“Xiao Hu”, with the model directing actions and the execution environment allowing it to actually read files, run programs, and write results.

This is a worthwhile approach to emulate. Although we don’t need Xiao Hu yet, our ownOffice Agent combined with the Ling-3.1-flash combo should also deliver good results.

As tasks continue to progress, the data and historical context will expand significantly, and the computational overhead will increase accordingly.

At this point, Ling-3.1-flash’s hybrid architecture design plays a crucial role.

The model has about 560B total parameters, with roughly25B activations per token. While preserving knowledge capacity, Ling-3.1-flash reduces the scale of weights that actually participate in computation per step.

In terms of attention architecture, it continues the Ling-3.0-flash design and further increases the proportion oflinear attention.

It uses a configuration of 7 layers of KDA plus 1 layer of Gated MLA, where KDA handles linear‑time sequence processing, and every 8 layers retain one Gated MLA layer to maintain a global attention pathway.

The model’s context window can reach up to1M, allowing it to accommodate more documents, code, and historical tasks in a single pass, reducing the effort of manual splitting and repeatedly moving information.

At the same time, the model configures 512 routing experts, selecting 8 of them for each token, and also invoking 1 shared expert.

Thismixed dense-and-sparse MoE approach has a single purpose: to keep the model from breaking down during long executions while avoiding excessive computational load.

In addition, the team is actively lowering deployment barriers; through mixed‑precision quantization, model weight usage has been reduced from roughly 534 GB to 292 GB, and in the corresponding Hopper test configuration, the required GPUs for deployment have dropped from 8 cards to 4 cards.

For specific business scenarios, the team also offers long‑context parallel inference solutions (PCP/DCP), and has enhanced support for JSON Schema, making the model more convenient for handling extensive data and integrating with existing business interfaces.

Beyond the architecture, whether it’s a mule or a horse, it still needs to be run through real evaluations.

In the office task benchmark GDPVal-AA V2.1, Ling-3.1-flash achieved an ELO score of 1673, demonstrating its competitiveness in handling multi-step work tasks.

△Image sourced from Bailing official WeChat

Meanwhile, according to the latest public benchmark by Artificial Analysis, Ling-3.1-flash has an intelligence index of 41, an output speed of about 210 tokens per second, and a mixed-call price of about $0.19 per million tokens, offering a choice that balances intelligence level, generation efficiency, and cost for high-frequency business scenarios.

Medical Scenarios In the HealthBench Professional evaluation under medical scenarios, Ling-3.1-flash also remains relatively ahead.

△Image source: Baoling official WeChat

In addition, in scenarios such as software development and terminal operations, Ling-3.1-flash also possesses a certain level of information comprehension and execution capability.

Take a long‑term Lua native compiler development task within Ant as an example; the entire process lasted about 17 hours, during which the model was iteratively developed and refined, ultimately achieving an independent test pass rate of 97.8%.

Therefore, when selecting an office Agent, Ling-3.1-flash’s comprehensive and error‑free professionalism makes it the top choice for trial deployment.

Real‑World Test of Ling‑3.1‑flash Shows Its True Power

To see how truly “real” it works, we ran a hands‑on test.

For the first challenge, we start with a Double‑11 e‑commerce task to dip our toes in.

Assume you are the e‑commerce head of the home organization brand Ling. Based on the data, you need to submit to your boss a Double‑11 promotion plan covering October 15 to November 11, which will be handed directly to operations, content, customer service, and warehousing for implementation.

P.S. The related data and materials were generated with the assistance of Codex.

Given the business objectives, product information, team situation, and historical channel data, the model quickly organizes the information and designs a major promotion plan.

First, the model conducts an initial target feasibility assessment.

It does not simply accept the boss’s lofty promises; by proactively identifying constraints such as diminishing marginal advertising efficiency, inventory, and shipping capacity, it provides a relatively reasonable estimate for the initial target.

It also provides the corresponding adjustment path to help workers cope with their boss’s inquiries.

Looking again at the promotional discount plan provided by the model, the key figures calculated actually hold up to verification.

Whether it’s the product bundle table or the channel budget details, every data point is traceable, indicating a certain level of data analysis capability.

Ling-3.1-flash also provides a detailed phased execution schedule, even including the prerequisites required for each phase.

Although the prompt only asked for the promotion plan up to Double Eleven day, the model independently added the subsequent review and profit calculation process.

It’s really warm and thoughtful~

The model’s text output is also quite good, for example it can handle various short‑video traffic‑driving prompts and live‑broadcast scripts with ease.

The visuals, copy, and hooks are all ready; we just copy the AI’s work directly, hassle‑free and convenient.

Finally, the model will proactively perform consistency checks, clearly indicating which data still need user verification and which content requires user confirmation, so the control over the final plan’s execution always remains with the user.

In addition to everyday office work, medical and health scenarios are also a strength of Ling-3.1-flash.

We provide a set of virtual lipid test data and vague symptom descriptions, hoping the model can generate a communication sheet to take to the clinic.

When information is insufficient, Ling-3.1-flash does not make an immediate judgment; instead, it first lists the details that need the user’s input, then distinguishes known facts, uncertainties, and questions for the doctor.

After guiding the user to provide key information, it cross‑checks the new content with the original content to produce a more professional consultation reference.

It is worth noting that, during its reasoning process, the model consistently emphasizes the red‑line warning principle, constantly reminding users to monitor changes in their condition and seek medical attention promptly, and it does not provide explicit medication guidelines.

In the process, it is more like performing preliminary assistance for doctors.

However, Ling-3.1-flash also has some minor bugs, such as when you feed the model a large block of text all at once,the model’s attention can sometimes drift.

For example, a Double‑Eleven promotion plan is inherently a very complex and long‑term task, so it’s better to deliver it incrementally—first organize the data with scripts, then generate the plan, proceeding step by step.

A single‑sentence generation also has its limits, and if the current result is unsatisfactory, it’s often more efficient for the user to step back and clarify the implementation approach and delivery specifications rather than keep guessing.

Moreoverthe current version does not yet have native visual understanding, requiring you to prepend a visual model yourself, which also increases the overall process cost.

In short, whether the model can achieve usable results varies depending on the task.

Bailey Leads the Way into Real-World Task Flows

After discussing the model itself, looking back atBailey, it actually hits the industry’s pain points precisely.

This year, the large‑model community has been almost entirely focused on coding and leaderboard competitions.

But after a certain point, more and more people are asking a more fundamental question: Is coding the only exam for AGI?

Xi Ting, head of Bailing, judged at the Bund Conference that Coding is not the only exam.

In his view, Bailing’s goal is still to serve real‑world tasks, extracting model capabilities from the technical layer and using real tasks to score the model.

Therefore, Baoling placed the 3.0 series models in the two areas where mistakes are least tolerable:Finance and Healthcare.

LingFin is a high‑efficiency financial large model that reconstructs financial productivity with industry research as the entry point; the healthcare‑enhanced model LingSante focuses on medical knowledge reasoning, evidence‑based retrieval, and standard workflow exploration.

Among them, Ling-3.0-flash-Fin (total parameters 124B, activation 5.1B) enables AI to enter real investment research workflows, connecting the full task chain from data gathering and analysis to producing research outcomes.

It rivals cutting‑edge 3‑trillion‑parameter models in high‑difficulty investment research tasks, with the financial Agent achieving open‑source leading performance and key task results comparable to frontier models such as GPT‑5.4.

Ling‑3.0‑flash‑Sante brings a high efficiency‑to‑effectiveness ratio to the medical domain, achieving the best performance among models under 1.6 T parameters with a 124 B parameter size,matching the overall performance of several flagship models.

△Image source: Bailing official WeChat

Compared to the baseline, these are not simple industry fine‑tuning; they are truly placed into complex work environments and continuously refined and validated.

From a broader perspective, Baoling is also doing something industry‑leading: redefining the term “efficiency”.

In the past, the industry talked about the compute‑efficiency ratio, how much compute was used and whether efficiency was high.

Baoling is ahead of the curve by tying together model compute efficiency, parameter efficiency, and token efficiency, emphasizing the model’s intelligence‑efficiency ratio.

This claim quickly received industry response, and Anthropic and OpenAI have both been rolling out professional capabilities for finance and healthcare over the past six months.

For example, Anthropic launched last monthClaude for Financial Advisors, organizing connectors and task Skills around specific jobs, which is another practice of integrating professional workflows.

Admittedly, the debate over approaches is still far from a clear winner, but “Entering the Real Workflow” is shifting from a single company’s choice to an industry-wide trend.

This year is also a pivotal year for Baoling as it moves from model capabilities to task value.

Through Ling-3.1-flash, the next steps for Bailing are already outlined:

Upward, continuing to raise the ceiling of Agent capabilities for developers;Downward, further lowering the usage barrier for ordinary users, turning a better intelligence‑efficiency ratio into tangible task‑handling ability;Outward, deepening model specialization to achieve breakthroughs in the real world.

Thus, the way the model community proves itself is shifting gears, but everything is just beginning.

Ling Studio：
https://ling.tbox.cn
OpenRouter：
https://openrouter.ai/inclusionai/ling-3.1-flash
HuggingFace：
https://huggingface.co/inclusionAI/Ling-3.1-flash
ModelScope：
https://modelscope.cn/models/inclusionAI/Ling-3.1-flash

One-click Triple Action‘Like’ ‘Share’ ‘Heart’

Feel free to leave your thoughts in the comments!

— End —

🌟 Light Up the Star 🌟

Daily Updates on Cutting-Edge Tech