E223｜The Year Applications Took Off: Model Technology Evolution and Commercialization
Summary
- The 2025 application boom was not driven by a breakthrough in any single model; it came from concurrent gains in reasoning, instruction following and tool use that pushed AI past the “good enough” threshold. 徐栋 traces the inflection point to o1 in late 2024: models moved from relying on fixed workflows to balancing generalization and accuracy on their own given sufficient context; standards such as Claude Skills and MCP also made tools explicit, moving AI from chat windows into SaaS, hardware and production workflows.
- Video generation is the modality with the clearest commercial loop today, with unit economics already strong enough to support short-form dramas, animated dramas and high-volume advertising. 徐栋 sees its capabilities as having moved from “GPT-3.5 to GPT-4”; a rational cost for an AI short drama can be pushed below RMB20K, while generating a 15-second ad can cost less than RMB10–15 against market delivery prices of roughly RMB25–50. With Qwen and Wanxiang combined into a pipeline, “5–6 people, or 10 people” can generate roughly 6,000 ads a day.
- Enterprise procurement has shifted from validating results to scrutinizing performance, concurrency and cost in detail—a sign that models are entering serious production. Customers now specify separate TPS and time-to-first-packet targets for 4K and 100K input; intelligent outbound calling may even require 100K calls to be completed within half an hour. Repetitive tasks, intent understanding and extraction are shifting to smaller models: “There’s no need to run a very large model.”
- Inference costs are still falling rapidly, with 徐栋’s rule of thumb being a decline approaching one order of magnitude every 6 months. Sparse architectures, mixed precision, MTP, model-chip co-design, cloud scheduling and overnight batch runs all contribute to lower costs; Qwen3-Next demonstrates a high sparsity ratio with 80B parameters and 3B active, while the on-device Omni 4B makes it possible to handle roughly 70% of general interaction, voice and vision-understanding tasks locally.
- Open source is not opposed to commercialization: Qwen uses community reach to pull demand into the cloud, then sells APIs through continuous updates, concurrency and large-model capabilities. “The moment it is open-sourced, it becomes an offline version”; closed-source services can still correct problems. Qwen3-Max was not open-sourced because of deployment costs, while DeepSeek’s release energized the open-source ecosystem and broadened adoption across Qwen’s 0.5B-to-235B range.
- The durable application-layer moat is not a simple wrapper, but the integration of proprietary data, models and workflows with industry decision-making. Insta360 uses an in-house model to understand scarce 360-degree footage before handing off to general-purpose models for orchestration, yet automatically editing 3 minutes of panoramic footage still costs more than RMB10. In a high-end consumer-electronics case, 语义科技 delivered 23% sales growth and more than 6% higher average order value after 18 months, then expanded from one business unit to the broader group.
- Token volume is useful for measuring current usage intensity but insufficient to represent long-term commercial value; revenue may ultimately come from Agent as a Service. The smallest and largest models may differ by 1000x in cost and revenue, while multimodal token definitions remain inconsistent. Model R&D spending has not declined; the next goal is for agents to “run continuously, say, for 2 weeks,” interacting with environments, using tools and iterating on hypotheses to deliver complete results.
Deep dive
1. Once model capabilities became “good enough,” the debate shifted from defects to the commercial loop
泓君 looks back 2 years to when the industry was still debating why large models struggled to enter work and daily life. Today, most phones have 1 or more AI assistants, traditional enterprises are deploying them at scale, and the question has become how technical evolution and commercialization can reinforce each other.
徐栋 divides Alibaba’s foundation-model portfolio into 3 lines: Qwen for language intelligence and production tool calling; Wanxiang for images, video and editing, with recent focus on 2.5 Preview and 2.6; and the pure-audio model Fun-Audio for ASR, TTS, voice cloning, multilingual support, dialects and accents.
His customer mix has also changed. Manufacturing, physical brands and internet companies remain the core base, but the past year has brought a wave of AI-native companies that have generated “some very good ROI” in fragmented markets.
2. Video generation is going through its own GPT-4 moment
徐栋’s clearest view is that, by analogy with language models’ move from GPT-3.5 to GPT-4, video generation “has already reached the GPT-4 state.” It is no longer mainly used for visual effects and entertainment, but is entering the production of animated dramas, short-form dramas and advertising.
Duration is the first technical axis. The once-common 5-second and 10-second clips are moving toward 15 seconds, with 1 minute potentially next. 徐栋 expects the end state to be 3-minute short videos generated directly from a few images or a single sentence, fundamentally changing the product experience.
The second axis is cinematography. Models can already switch between multiple shots and change lighting, compressing capabilities that once required coordination among directors, cinematographers and art departments into “professional, film-grade” capabilities that ordinary users can invoke through prompts.
3. Character consistency is moving video creation from “抽卡” to controlled production
徐栋 describes the key lesson from Sora 2 as character consistency: a user records roughly 5 seconds of selfie video, turns around and says a few lines, and the model may preserve the person’s appearance and voice timbre for continued reproduction in later creations.
泓君 asks whether this means a real person only needs to appear briefly before AI takes over. 徐栋 confirms that the core change is greater controllability in generated content. “We may have previously relied on抽卡”; now users can add more references at the input stage.
Anime production has already moved from reference images to reference video, because video contains more information about the form and size of people, objects and backgrounds, as well as the character’s voice timbre. 徐栋 believes character consistency will ultimately become standard in video-generation models.
4. Short-form dramas and advertising have already produced calculable unit economics
In 徐栋’s view, China’s short-form drama market has surpassed the film market, with content production partially shifting from live-action shooting and heavy manual editing to AI. The fastest-growing animated dramas began with dynamic comics that add motion to still images, offering relatively strong serialized storytelling and commercialization potential.
With quality tiers such as S, A and B, a rational cost for an AI short drama may be below RMB20K. Combined with traffic acquisition, that can be enough to break even or generate a decent return; premium productions still require additional post-production resources, with cost rising alongside finished quality.
The AI generation cost for a 15-second ad can be kept below RMB10 or RMB15, while qualified assets command roughly RMB25–50 in the market. A pipeline built on Qwen and Wanxiang could allow 5–10 people to produce roughly 6,000 ads a day for agencies, advertisers or traffic platforms.
5. Language-model progress is becoming incremental, but directly determines speed and cost
As high-quality data grows scarce, model companies are turning to “fine craftsmanship”: restructuring data from different angles and changing its ordering to improve the efficiency of knowledge learning. 徐栋 believes increasingly impressive performance on corner cases mainly comes from this data engineering.
Architectures are becoming more sparse, while techniques such as MTP enable prediction of multiple tokens at once. 徐栋 gives a potential path for time-to-first-packet to fall from 2 seconds to 500 milliseconds, depending on context length; TPS could rise from a solid 30–50 previously to 80–100 or more, which he says is “very likely.”
Higher speed means greater machine throughput and lower costs. At the same time, teams continue to refine instruction following, context length, coding and tool calling. The challenge is not merely using one tool, but using tools in a generalized way, calling multiple tools continuously and executing tasks in depth.
6. The 2025 Agent boom was triggered by 4 capabilities arriving together
The first was reasoning. 徐栋 traces the fundamental inflection point to o1 in the latter half of 2024, close to year-end: models stopped merely relying on so-called probabilities and began to exhibit a degree of logical preference.
The second was controllability at larger scale. Previously, workflows had to hard-code key nodes and cross-check model bias. Now, as complex-instruction understanding and adherence improve, a model with sufficiently accurate context can find its own balance between generalization and accuracy.
The third and fourth were tool use and open standards. Claude Skills, MCP and similar standards have gradually gained acceptance, turning tools from implicit capabilities into callable interfaces. Models thus “broke out of their original input-output window” and entered SaaS workflows, production processes and even hardware interactions.
泓君 asks whether instruction following comes from some proprietary technology. 徐栋 is more measured: it is mainly about datasets and context understanding. “There aren’t actually that many secrets left in model training today”; every company is pursuing the same kind of fine-grained optimization.
7. The real novelty in AI hardware is completing an action, not adding another recognition feature
ASR, CV and hardware have been combined for 10 years, 徐栋 says. What is different this time is that models are more humanlike: they can listen, see, answer and execute tasks, rather than merely recognizing signals captured by a camera, screen or microphone.
He uses ordering coffee through the Qwen App as an example. A user selects the brand, flavor, sugar level and delivery time in natural language, then completes payment. The system can recommend options based on distance, preferences and history. On glasses, the ideal experience would not copy an App menu, but use memory to proactively shorten the interaction.
泓君 frames the input as the glasses seeing a menu or a colleague’s drink, and the output as voice. 徐栋 adds that the true output is an “action” enabled by integrating the commercial service. That is why AI Phone, AI PC, smart cockpits and similar products are redefining their positioning.
8. Alibaba’s commercialization strategy has 2 main tracks: enterprise productivity and product experience
Unlike cloud computing, which gradually spread from mobile gaming into finance and manufacturing, large models have covered every industry from the start, with differences mainly in penetration. 徐栋 therefore sees commercialization less in terms of industries and more in terms of 2 combinations: models with enterprises, and models with products.
Inside enterprises, the value lies in rebuilding workflows around outcomes across design, coding, research, sales and other roles. User-facing Apps, phones, cars and consumer electronics use models to change the interaction model, and in turn redefine the product itself.
For phones and cars, low latency, local processing and privacy needs are driving edge-cloud integration. Alibaba’s Omni 4B targets relatively powerful devices but still has to manage power consumption and heat. 徐栋 says the architecture has been refined for roughly 2–3 years and is now entering the “deep water.”
He estimates that “possibly more than 70%” of general interaction and basic voice and vision understanding can be handled on-device. Enterprises can also connect devices and business systems through Alibaba’s Bailian MaaS platform, with high concurrency, low latency, observability and governance.
9. Customers have shifted from asking “does it work?” to buying production-grade performance
Early customers built large numbers of vertical evaluation sets to validate results first. Over the past 6–12 months, their attention has clearly shifted to performance and cost. 徐栋 sees this as evidence that models are entering serious production after their effectiveness has largely been validated.
Requirements have become highly specific. Customers separately set TPS and time-to-first-packet targets for 4K input and 100K input. Intelligent outbound calling, whose effective window is concentrated, may require 100K calls to be completed within 30 minutes, making model concurrency directly relevant to business feasibility.
Demand for small models has been especially strong over the past 3 months. Repetitive tasks do not require the largest model, while first-step intent understanding and extraction for chatbots or agents also suit smaller models. Customers are beginning to combine closed-source data with data generated by closed-source models for post-training or fine-tuning.
10. Falling inference costs reflect the combined effect of models, chips and cloud scheduling
徐栋’s industry rule of thumb is that inference costs are “basically declining at a rate approaching an order of magnitude every 6 months.” An order of magnitude means roughly 10x, and the decline is not coming from model algorithms alone.
On the model side, the drivers are sparse architectures, mixed precision and MTP; on the chip side, vertical co-design with models. The cloud must also improve utilization through model switching, time-based task scheduling and overnight batch runs, avoiding higher unit costs when compute utilization sits at only 30%–50%.
Qwen3-Next is a concrete example: it has 80B parameters and 3B active, with a higher sparsity ratio than the earlier 30B-A3B model. 徐栋 says developers can already observe meaningful changes in inference efficiency, quality and speed.
泓君 uses Google TPU’s software-hardware coordination as a comparison and argues that Alibaba’s ownership of cloud, models and chips creates a three-way integration. 徐栋 adds that vertical integration is not only about chips and models; scheduling gains also come from the cloud.
11. Open source is the customer-acquisition layer for cloud commercialization, not a free substitute
徐栋 recalls that around August 2023, community feedback on Qwen 7B opened up the global landscape. Open source attracted leading researchers and students already familiar with models, while also allowing developers of inference frameworks, inference engines and on-device deployments to discover bad cases early.
Qwen spans 0.5B to 235B, allowing enterprises to choose by use case. The 1.5B and other sub-3B models suit small deployments, while 32B is a “sweet-spot model” because it can run on a single GPU.
Open source creates a virtuous cycle for Alibaba Cloud, but “the moment it is open-sourced, it becomes an offline version” （开出那一个刹那，它就是一个离线的版本了）. APIs can continuously correct problems while offering higher concurrency and an easier environment. Qwen3-Max was not open-sourced because deployment costs are high, and workloads Qwen3-235B cannot satisfy can shift to Max as a service.
Asked whether DeepSeek caused a shock, 徐栋 gives the opposite answer: its paper advanced the technology, and its release energized the entire open-source ecosystem. Qwen’s broad range of model sizes and high-frequency updates make the 2 sides look more mutually reinforcing than competitive.
12. Tokens can measure usage intensity, but cannot replace revenue or product value
徐栋 acknowledges that token volume is a good current indicator of usage and can temporarily inform startup valuations. Over the long term, however, the market will return to revenue, which may not come directly from tokens but from agents built on top of them.
The metric has clear shortcomings: the smallest and largest models may differ by 1000x in cost and revenue, while multimodal tokens for voice, text and other modalities lack a unified definition and vary slightly across companies.
Model as a Service focuses on model updates, API availability, realtime capabilities and batch runs, resembling the evolution from Assistants API to Responses API. Agent as a Service calls multiple models and packages standardized task outcomes; Deep Research is 徐栋’s example.
13. Model R&D spending has not receded; breakthroughs are simply less visible
Asked whether the industry is now more focused on commercialization, 徐栋 offers a contrarian view: “Model R&D investment has actually increased.” Every company and lab says it is “not pushing hard enough,” suggesting that competitive outcomes can still be measured quantitatively.
The difference is that the industry is seeing fewer immediately visible paradigm shifts like o1 reasoning or Nano Banana editing. More spending is going into data cleaning, evaluation standards and engineering details; even rule-based cleaning and evaluation that used to be done manually can now be performed by models.
The agentic capability he wants is for models to “run continuously, say, for 2 weeks,” continually interacting with the environment, thinking independently, making judgments and forming hypotheses. Only by combining tool use, long context, complex instructions and reasoning can models deliver research or business outcomes better than simple text in and text out.
泓君 summarizes this as painstaking optimization that may not qualify as an intelligence leap. 徐栋 leaves room in the definition: if intelligence can be scored, these engineering improvements can still keep raising the intelligence score.
14. Insta360’s editing problem is not stitching clips together, but correctly inferring creative intent
齐鲁教授 says Insta360 wants to automatically turn footage accumulated over a day—or even a year—into finished videos users are willing to share. Cloud-based automatic editing is already live and is a key step in the company’s strategy of “recording and sharing life better.”
“The real difficulty is not stitching the user’s clips together, but hitting the user’s editing intent.” Some users want a documentary-style continuous story; others want only highlights and a fast, satisfying rhythm. Still others care more about the emotional tone of the music, cinematography and transition style.
The collaboration has 2 layers: Qwen’s multimodal model handles perception, understanding, judgments about people and events, and structured orchestration; the Wanxiang series handles video-effects generation. To identify failed skiing moments, for example, the model must understand the instruction and recognize the actual failed action in the footage.
Semantic understanding now meets some basic needs. The harder problem is that the intent may not exist explicitly in the footage, and users may not know how to express it. 齐鲁教授 agrees with 泓君’s analogy: if users must act like a “good client” and precisely define a shot of someone flying off a sled, the experience has already failed.
15. 360-degree footage is Insta360’s proprietary-model moat, but it also raises costs
Panoramic cameras follow a “shoot first, edit later” model: capture the full 360-degree scene, then choose the best perspective, pacing and camera movement during editing. Even relatively uniform scenes can be made more dynamic through camera-motion effects. Compared with the huge volume of flat footage, general-purpose models are weaker at understanding 360-degree material.
Insta360 has developed its own panoramic multimodal model. It first extracts highlights from panoramic footage, then hands off to general-purpose models for planar content understanding and orchestration. 泓君 compares this capability to a world model, and 齐鲁教授 agrees.
World-model development still depends on real-world data, continuously expanding training methods and stable cloud compute. The VR drone shown at CES can also generate panoramic data, but 齐鲁教授 emphasizes that data enters model development only with user permission and compliance requirements satisfied.
The commercialization bottleneck is cost: processing roughly 3 minutes of 360-degree footage for a user may currently require RMB10-plus. Insta360 hopes to exploit the relatively high compressibility of panoramic footage, first reducing model input and token counts, then lowering automatic-editing costs.
16. The end state of enterprise AI is shifting from saving 5 people to finding scalable demand
吕英杰 says 语义科技 provides consumer-insights SaaS to brands including Nestlé, Procter & Gamble and Pop Mart. After the traffic dividend peaked in 2023, brands moved from simply buying traffic to combining traffic with consumer thinking, turning customer-service conversations into analyzable intent data.
Customers that directly use general-purpose large models to label raw corpora often achieve accuracy below 70%. Since 2021–2022, the team has trained vertical small models to classify information into consumer profiles, product and service feedback, and internal attribution. At one consumer-goods company, responsibility assignment for returns fell from 6 people recording cases to 1 person reviewing them.
吕英杰 does not believe AI will fully replace customer service. After knowledge-base calibration, single-point product Q&A can be more accurate than a human; but accuracy still falls on continuous, complex questions involving coupon stacking with regional subsidies or A/B basket-building. The foundation model must also handle intent recognition, RAG retrieval, and the cleaning and organization of complex knowledge sources including detail-page images and Chain workflows.
From the second half of 2024, customers shifted their focus from cost reduction to productivity gains. In a high-end home-appliance category with an average order value above RMB20K, AI VOC profiles were used to redesign selling points, livestream-host scripts and customer-service messaging. After 18 months, sales were up 23% and average order value had risen by more than 6%; in 2025, the approach expanded across the group’s business units.
Asked whether the dividend disappears once everyone has AI, 吕英杰 says “the core of AI’s future value will still be controlled by people.” A vertical data flywheel and industry experts form the middle-layer moat; humans will ultimately identify opportunities and decide on products, marketing and execution.
Its pricing combines a SaaS license with data-usage packages, estimating token costs for units such as 1 million conversations, voice interactions or social-media posts. AI has made customers more willing to accept that usage necessarily creates costs, improving the payment logic of China’s To B SaaS market. 语义科技 therefore achieved roughly 100% growth over the previous year.