AI Trends That Built Up Over Spring Festival, Before I'm Back at Work

Back to work tomorrow. Let me wrap up a few trends that built up in the AI world over this Spring Festival.

Honestly, I spent this Spring Festival in Singapore with my kid, phone glued to my hand. My smart avatar, OpenClaw, kept popping notifications: GLM-5 is out, MiniMax M2.5 is open-sourced, Doubao is on the CNY Gala, Qwen is treating the whole country to milk tea, and Yuanbao just dropped 1 billion yuan in red packets. I was putting my kid down for a nap while scrolling until 2 a.m., and the more I read, the clearer it became—these people aren't on vacation. They're grabbing territory.

My core thesis is one sentence: code keeps depreciating, while entry points and verification keep appreciating. Whoever can convert "traffic" into "reusable engineering capability" wins the next phase.

Cover image prompt (baoyu-cover-image, strict 2.35:1): Type=conceptual, Palette=warm, Rendering=digital, Text=title-subtitle, Mood=bold. Scene: left side shows a Spring Festival city nightscape with phone red-packet pop-ups; right side shows a data center with a model parameter panel; a glowing runway in the middle connects "Traffic Entry" and "Engineering System." Three abstract labels are distributed across the top: "New Model Launch," "User Acquisition," and "Engineering Rebuild." Title: "The Spring Festival AI War: Red Packets Up Front, Models Behind." Subtitle: "Day One Back: The New Coordinate System Every Programmer Must Rebuild." Requirements: clear text, no large English lettering, techy feel with a Spring Festival atmosphere, Aspect Ratio 2.35:1, output PNG.

If you're short on time, just remember three things:

• This Spring Festival push goes far beyond "sending red packets." It's a battle for AI entry points—total investment exceeds 4.5 billion yuan, and the core battlefield isn't the red-packet amount but who locks in that "ask AI first" slot on your phone.
• The real deciding factor lies beyond the marketing budget: whether model capability can be baked into workflows. GLM-5 and MiniMax M2.5 shipped in the same week, and the signal is clear—model capability is the foundation of retention.
• For programmers, the first day back isn't about swapping tools; it's about redefining delivery. Your value is no longer measured by "how many lines of code you wrote."

Let's Lay Out the Facts First

I like to list verifiable public information before making judgments. The data below comes from each company's official website, official social media, and public reporting.

Model Side: Two Flagships Collide in the Same Week

One thing I noticed: GLM-5 and MiniMax M2.5 were released in essentially the same week. This is no coincidence. Shipping a model during the Spring Festival window is about seizing narrative control during the "everyone's paying attention to AI" period—while you're watching the Gala, the model has already quietly moved to the next generation.

GLM-5's narrative arc is clear: a shift from "Vibe Coding" to "Agentic Engineering." Zhipu put that phrase on the title page of its technical report, meaning AI is no longer just helping you write code—it's participating in the full engineering pipeline as an autonomous agent. Another signal I read twice: "full-pipeline training on domestic chips"—the entire process from pre-training to alignment was completed on Huawei Ascend, which carries special weight at the geopolitical level.

MiniMax M2.5 takes a different path: it doesn't chase parameter scale; it chases "usability." 230 billion total parameters with only 10 billion activated at inference—extremely fast, extremely cheap. Less than a day after launch, global users had built over 10,000 experts on MiniMax Agent. I tested the Agent's response speed myself, and the feel is genuinely in a different league from models I'd used before. The data says it: speed and cost advantages translate directly into user engagement.

Illustration 1 prompt (baoyu-article-illustrator, strict 2.35:1): Type=timeline, Style=editorial. Title: "2026 Spring Festival AI Major Events Timeline." Content: a timeline showing "2/11 GLM-5 Released (744B parameters)" → "2/12 MiniMax M2.5 Launched" → "2/13 M2.5 Open-Sourced Globally" → "New Year's Eve: Doubao CNY Gala AI interactions hit 1.9 billion" → "During Spring Festival: Qwen's 3-billion-yuan treat / Yuanbao's 1-billion-yuan red packets" → "2/22 First Day Back"; each node color-coded to distinguish "Model Events (blue)" from "Entry-Point Events (red)." Requirements: timeline infographic, clear nodes, readability first, Aspect Ratio 2.35:1.

Entry-Point Side: Red Packets Are Really "Try AI Once"

According to reports, total investment in this red-packet war exceeds 4.5 billion yuan.

Let me draw a boundary: the exact figures each company invested vary across media outlets (some say 4.5 billion total, others say 8 billion). This article uses the data points on which sources largely agree.

Beyond the numbers, I focused on three things:

• Doubao secured a "Gala-level entry point." The reach of the CNY Gala needs no explanation. ByteDance used that entry point to get hundreds of millions of people interacting with AI for the first time in a holiday context—not search, not Q&A, but "send red packets + generate avatars + write blessings," a scenario everyone can join. On New Year's Eve alone, AI interactions hit 1.9 billion and 50 million avatars were generated. I sent my mom an AI-generated avatar, and she actually kept it for three days without changing it. Consumer-side AI usage habits are being built at high speed.
• Qwen is stitching together an "e-commerce + local services" chain. It's not just "treating you to milk tea"—it's connecting Qwen to Taobao Flash Buy, Fliggy, Hema, and other Alibaba ecosystem businesses. You claim coupons, order food, and place orders inside Qwen, and all the data flows back into the Alibaba ecosystem. This is the textbook play of turning an "AI assistant" into a "consumption entry point."
• Yuanbao is walking the well-trodden "social virality" path. Tencent executives openly said they want to "recreate the success of WeChat red packets back in the day." Leveraging the WeChat and QQ ecosystems for share-and-spread virality, it uses the simplest "send red packet—grab red packet" mechanic to pull users into the Yuanbao app.

Why I Call This an "Entry-Point Rebuild War"

A lot of people see this as "holiday marketing." That's only the surface layer I'm seeing.

The surface is red packets. Beneath it, three things are shifting:

• Entry-point reshuffling: whoever first claims the "open this app when you hit a problem" slot on the user's phone gets the interaction data and usage habits first.
• Mindset reshuffling: users' positioning of AI is upgrading from "search replacement" to "task agent"—not just looking up information, but doing things for me, placing orders for me, writing things for me.
• Workflow reshuffling: on the enterprise side, model capability is being embedded into real business pipelines. Behind the "milk tea treat" you see is Qwen's API integration with Taobao's order system; behind the "Gala interaction" you see is Doubao's large-model real-time inference cluster under load.

Illustration 2 prompt (baoyu-article-illustrator, strict 2.35:1): Type=comparison, Style=blueprint. Title: "The Spring Festival AI War: Surface Red-Packet Battle vs. Underlying Engineering Battle." Content: left side "Red-Packet Battle (Surface)" includes 4.5B+ in acquisition subsidies, Gala exposure, social virality, and topic diffusion; right side "Engineering Battle (Underlying)" includes GLM-5/M2.5 model capability, Agent task success rate, API ecosystem integration, and user retention & repeat usage; a thick arrow in the middle labeled "Traffic → Usage Habit → Workflow → Revenue Loop." Requirements: comparison diagram, engineering blueprint style, high-contrast key information, Aspect Ratio 2.35:1.

In other words: red packets are the "door-opening move." Models and engineering are the "compounding move."

Here's something from my own experience. With Qwen's "milk tea treat" this time, after I claimed a 25-yuan free-order coupon, the actual order path was: Qwen → Taobao Flash Buy → merchant → delivery. The entire chain worked end to end, which means Qwen has gone from a "chat box" to an entry point that can trigger real transactions. If this entry point retains users, Alibaba gains a new traffic distribution channel.

Same logic: Doubao's "AI avatar generation" and "AI blessing generation" in the Gala context are essentially implanting a perception in users: there's an AI app on your phone that can do fun things for you. Once that habit is established, downstream commercialization scenarios (writing assistance, image generation, video creation) have a foundation.

The Real Model-Side Showdown: It's About "Who's More Usable"

I'm putting GLM-5 and MiniMax M2.5 side by side because they happen to represent two typical routes in today's Chinese AI model competition. I've hit the API quirks of both, so here are some hands-on impressions.

GLM-5: Ascend Chips + Full Parameter Scale + Agentic Engineering

In GLM-5's technical report, a few points caught my eye:

• 744 billion total parameters, 40–44 billion activated at inference—top-tier among open-source models at this scale.
• 200K context window with integrated DeepSeek Sparse Attention (DSA). Long-context capability is preserved, and inference cost stays in check.
• Full-pipeline training on Huawei Ascend chips. Prior large-model training was almost entirely on NVIDIA GPUs; this time it's explicitly labeled "All-Ascend Pre-training," a significant signal for technical self-sufficiency.
• SWE-Bench Verified: open-source SOTA. The official line on real-world coding experience is "approaching Claude Opus 4.5."

Zhipu used an interesting positioning phrase: "From Vibe Coding to Agentic Engineering"—AI shouldn't just "help you spit out code fast" (Vibe Coding); it should act as an autonomous role within the engineering team, understanding tasks, breaking down requirements, executing automatically, and self-verifying. I ran a medium-complexity refactoring task, and GLM-5 genuinely decomposed it into subtasks and executed them one by one. That behavioral pattern is noticeably different from models I'd used before.

MiniMax M2.5: Extreme Efficiency + Open-Source and Ready to Use

MiniMax M2.5 takes a completely different approach:

• 230 billion total parameters, only 10 billion activated at inference—deployment barrier and inference cost drop dramatically.
• Output speed of 100+ tokens/s, roughly 2× that of mainstream models. In Agent scenarios, speed is the experience.
• Input at roughly $0.3 per million tokens, output at roughly $2.4 per million tokens—very competitive for complex Agent applications.
• SWE-Bench Verified score of 80.2%; in Multi-SWE-Bench and other multilingual, complex environments, it beats Opus 4.6.
• Open-sourced globally the very next day. From release to open-source was under 24 hours; less than a day after launch, over 10,000 experts had been built.

The difference between the two routes is clear:

GLM-5 bets on "absolute capability"—parameters big enough, capability strong enough, to handle the most complex Agent tasks and long-horizon engineering pipelines. The implicit assumption: future Agents will need a sufficiently powerful base model to understand complex intent.

MiniMax M2.5 bets on "usability"—the model must be fast, cheap, and deployable immediately. The implicit assumption: users and developers care more about "how it feels to use," and paper parameters come second.

My take: both routes are valid. The edge goes to whoever first closes the "model → Agent → workflow → revenue" loop. Big parameters you can't actually use are wasted; high efficiency without enough capability can't hold up in complex scenarios. The deciding factor is the deployment data over the next few months.

What Does This Have to Do with Software Engineering? A Lot.

If you still treat AI as a "code-writing plugin," you'll underestimate this shift. Last year I still thought Copilot was enough. After getting my face slapped by an Agent workflow in January this year, my attitude changed completely.

What was AI Coding before? Copilot, auto-completion, "you write half, it fills in the other half"—the programmer is still the lead, AI is the supporting actor.

What is Agentic Engineering now? The Agent understands requirements, breaks down tasks, generates code, runs tests, submits PRs, and rolls back to retry when it hits a problem—the programmer's role is shifting from "implementer" to "orchestrator + reviewer."

For engineering teams, the real change has four layers:

• Requirements layer: users want "just give me the result." Suggestive answers are losing their pull. Qwen places your order, Doubao generates your avatar—users don't want to see the intermediate process, just the final output.
• Architecture layer: single-model strategies are giving way to multi-model routing. Flagship models handle complex tasks; lightweight models handle high-frequency, simple requests. You have to balance cost and experience on both ends.
• Quality layer: the share of manual review is dropping; the weight of automated verification is rising. Anthropic's data shows that roughly 60% of work already involves AI, but the proportion of tasks fully delegated to AI is still in the 0–20% range. Human oversight remains critical.
• Organization layer: new roles are emerging within teams—Agent Orchestrator, Prompt Engineer, AI QA Specialist. The traditional "product → dev → test" three-stage delivery model is being restructured.

Code itself is depreciating. Intent and tests are the assets being repriced.

Illustration 3 prompt (baoyu-article-illustrator, strict 2.35:1): Type=framework, Style=minimal-flat. Title: "2026 Agentic Engineering Operating Framework." Content: top layer "Business Intent Input" → middle layer with three parallel roles: "Architect Agent (understand goals / decompose tasks)," "Coder Agent (generate code / invoke tools)," "QA Agent (build tests / block substandard delivery)" → bottom layer "Deliverable + Automated Verification Report"; left sidebar "Multi-Model Routing: Flagship (complex tasks) / Lightweight (high-frequency requests)"; a circular arrow between QA and Coder labeled "Fail → Roll Back & Retry." Requirements: framework diagram, clear structure but professional, Aspect Ratio 2.35:1.

Day One Back: 5 Things You Must Figure Out First

Change Your KPIs First: Code Is Not the Final Deliverable

Starting today, the way a team measures output should shift to these three things:

• Is the intent clear? (Are the requirement boundaries well defined?)
• Is the verification complete? (Can the tests and acceptance criteria actually be executed?)
• Is the outcome traceable? (Can a failure be traced back to a specific stage?)

If your team is still using "lines of code" or "PR count" as the core metric, today is the day to change it. I was still complaining to a friend last week—their CTO was reviewing "per-capita code output" in the weekly meeting. I told him to cut that metric, period.

Treat AI as a "Team," Not Just a "Chat Box"

The minimum viable Agent Team is still this triangle:

• Architect Agent: understands business goals, decomposes tasks, defines interfaces.
• Coder Agent: implements code, invokes tools, iterates and commits.
• QA Agent: builds tests, runs regressions, blocks substandard delivery.

Both GLM-5 and MiniMax M2.5 are moving in the "Agentic" direction, which means this three-role collaboration is already baked into model capability—it's not just a concept anymore.

Set Hard Rules: Accept First, Generate Second

Write at least these four rules into your team template:

• Acceptance criteria are written before the Agent starts coding.
• Every generation must pass automated tests.
• If quality, performance, or security falls short on any dimension, the merge is blocked.
• Failure paths must be rollback-able, reproducible, and reviewable.

The Moat Has Changed: Focus on Four Capabilities

• System design capability—AI can write code, but it doesn't know what architecture to build.
• Test-driven capability—the quality of your acceptance criteria defines the ceiling of AI output quality.
• Multi-Agent orchestration capability—how do multiple Agents collaborate without stepping on each other.
• Prompt + Skill Engineering—distill experience into reusable, transferable "skill definitions."

Execute Immediately: Day-One Minimum Action List

• Pick 1 real requirement from this week. Don't build a demo.
• Write 10 actionable acceptance criteria.
• Wire up the Architect/Coder/QA three-role loop.
• Put tests, logging, and rollback conditions into the PR template.
• Evening retrospective: what headcount did you save, what new risks did you introduce.

Illustration 4 prompt (baoyu-article-illustrator, strict 2.35:1): Type=framework, Style=notion. Title: "Day-One Execution Checklist × Risk Dashboard." Content: left side "5-Item Minimum Action List" (pick a requirement, write acceptance criteria, build the loop, update the template, evening retrospective); right side "4-Dimension Risk Dashboard" (Quality Risk: AI hallucination / regression failure, Cost Risk: model API bills, Compliance Risk: data security / licensing, Stability Risk: Agent runaway / infinite loop); bottom section "Same-Day Retrospective Conclusion." Requirements: framework diagram, practical orientation, clear and readable, Aspect Ratio 2.35:1.

Looking a Step Further: 3 Trends After This Wave

This Spring Festival push is just the beginning. Looking ahead, three things will keep fermenting through 2026:

Multi-Agent Parallelism Becomes the Default

Most AI Coding tools today still run "one Agent, one task." But GLM-5's 200K context window and MiniMax M2.5's extreme inference speed are both paving the way for "multiple Agents working simultaneously." Gartner predicts that by 2026, AI agents will achieve task coherence spanning weeks. By then, you might have 5 Agents developing in parallel on different branches, and your job is review and merge.

The "Entry Point → Workflow → Revenue" Loop Competition Has Barely Started

Spring Festival red packets only solved the "get users to open the app" question. Two more hurdles sit behind it: keep users around (workflow value) and get users to pay (revenue conversion). Those two hurdles are the real competitive moats.

The "Appreciation Direction" for Programmers Is Clearer Now

The direction isn't "learn more languages" or "grind more LeetCode." It's:

• Designing system architectures that let Agents execute efficiently
• Defining acceptance criteria that make AI output verifiable
• Orchestrating multi-Agent collaboration without things falling apart
• Distilling experience into reusable Skills and Prompts

What's being repriced is "the precision with which you define the problem." "How fast you can type code" is the thing depreciating.

In Closing

The one thing this Spring Festival made most certain for me: whether AI belongs in the main pipeline is no longer the question—AI is already in the main pipeline. The question is who hasn't finished the organizational upgrade yet.

GLM-5 and MiniMax M2.5 shipping in the same week looks like a "parameter arms race" on the surface, but in substance it's a "fight for definitional power"—whoever first defines the standard workflow for Agentic Engineering holds the discourse power over the next generation of software engineering.

Doubao, Qwen, and Yuanbao scattering billions in red packets looks like "burning cash for acquisition" on the surface, but in substance it's a "fight for the entry point"—once users get used to "ask AI first," the downstream commercialization chain falls into place.

So what you really need to figure out on day one back is: is my team ready to shift from "writing code" to "orchestrating Agents"? "Which model to pick" is a later question.

If you're not ready yet, start with those 5 action items from today.

References (Official Sources)

• Z.AI Release Notes: New Released (incl. 2026-02-11 GLM-5) https://docs.z.ai/release-notes/new-released
• Z.AI Docs: GLM-5 Overview https://docs.z.ai/guides/llm/glm-5
• Zhipu AI Official: GLM-5 Technical Report (All-Ascend Pre-training, Agentic Engineering positioning) https://zhipuai.cn/
• MiniMax API Docs: Models Release Notes (incl. Feb. 2026 M2.5 release) https://platform.minimax.io/docs/release-notes/models
• MiniMax API Docs: Text Generation (M2.5 capability and efficiency metrics) https://platform.minimax.io/docs/guides/text-generation
• MiniMax M2.5 Hugging Face open-source page https://huggingface.co/MiniMaxAI
• Tongyi Official Site (Qwen / Wan capability portal) https://qianwen.aliyun.com/
• Doubao × 2026 CCTV CNY Gala Exclusive AI Cloud Partner https://research.doubao.com/
• Tencent Yuanbao Spring Festival Red Packet Campaign (official Weibo) https://www.weibo.com/7914117279/QrLY7xX5i
• Alibaba Qwen "3-Billion-Yuan Spring Festival Treat Plan" https://qianwen.aliyun.com/
• Jiemian News: 2026 Spring Festival AI Red Packet War – Comprehensive Report https://www.jiemian.com/
• Anthropic: Agentic Coding Trends 2026 (AI usage ratios and human oversight data) https://www.anthropic.com/