On August 13, 2026, DeepSeek completed a dual release with sharp industry signal, simultaneously launching its new flagship model V4 Pro and the open-source agent framework DeepSeek Harness. The former carries the core capabilities of model inference, logical reasoning, and solution decision-making — the brain of AI; the latter handles tool calling, device operation, task execution, and workflow orchestration — the body and hands that bring AI into the real world. V4 Pro adopts a Mixture-of-Experts architecture with 1.6 trillion total parameters, activating only 49 billion per inference, an expanded context window of one million tokens, and a maximum output approaching 380,000 tokens. On paper, its overall capability is already a credible match for the full flagship lineups of Claude and GPT.
But what truly shook the entire AI developer community and reshaped the industry's competitive logic was not the iteration of model parameters, but the simultaneously open-sourced Harness framework. Within 24 hours of launch, its GitHub stars surpassed 80,000, daily momentum far exceeding the long-accumulated totals of xAI's Grok series, and the companion plugin repository quickly passed 1,300 entries. That a developer-preview product could forge industry consensus so fast confirms a long-overlooked truth: the AI race no longer lacks models good at dialogue and reasoning. What is truly scarce is an execution runtime layer that embeds deeply into real workflows, is transparent and controllable, can be self-modified, and can actually deliver.
Set this launch against the closed-operating paths of earlier industry leaders, and the deepest fault line in today's AI industry comes into focus. Two growth logics coexist: one binds users through closed ecosystems, privacy lock-in, and high switching costs, building moats on information asymmetry and black-box rules; the other opens the underlying code entirely, returns control to users, and builds a public ecosystem, trading transparency and co-construction for long-term trust and industrial value. Readers may judge which is superior; this article only dissects both routes in parallel, clarifying the underlying logic, durable value, and eventual outcome of the two business models.
I. Harness: From "Can Think" to "Can Deliver"
For years, the entire industry's iteration focus has been almost entirely on the model's "brain power" — longer context, stronger logic, higher benchmark scores, smoother dialogue — the dimensions every vendor raced to out-spec the others on. But no matter how far a model's reasoning evolves, pure intelligent inference stays at the virtual layer, separated from real commercial and production deployment by an entire execution system. Device operations, file read/write, command execution, tool scheduling, multi-step task decomposition, workflow closure — these real production actions are the "bodily capability" most models simply lack. DeepSeek Harness was born precisely to fill this industry gap. Its core definition is blunt: an agent equals model capability plus the Harness execution layer. The model decomposes requirements, formulates plans, and reasons logically; Harness turns abstract plans into concrete computer operations and task execution, fully closing the last mile from AI thinking to delivery. Its direct benchmark is exactly the mainstream high-end office and coding agents — Anthropic's Claude Code and OpenAI's Codex — which means the industry's competitive center of gravity has shifted wholesale from single-model reasoning to the landing capability of the agent runtime layer. The most subversive engineering design of this framework is its extreme plugin-based, modular, decoupled architecture. Model adapters, tool registration, skill invocation, session management, sandbox protection, data storage, task scheduling, and the interaction UI — every core component is independently encapsulated as a replaceable plugin. Developers can freely swap, recombine, and customize functions to fit their own scenarios without touching the underlying source. Most critically, DeepSeek Harness is a fully model-agnostic open framework: it does not force-bind its own model, and is compatible with all mainstream large models, letting every model gain real working capability by relying on this mature execution system. The framework ships with four preset modes — general office, programmatic tool calling, lightweight file operations, and creative debugging — supports quick command-line launch and autonomous source-code deployment, and adapts to development and office needs at different levels. In real capability, it covers the full chain: code reading, file modification, command execution, full-network search, multi-stage task planning, and sub-agent dispatch; it enforces mandatory human approval for sensitive operations, fully retains tool-call logs and interaction records, supports session branching and step rollback, and carries the complete closed-loop maturity of an engineering-grade product — already well beyond a demo-tier product, it is a doer-grade runtime that can be deeply embedded into everyday R&D and office workflows. From the perspective of industry cycles, the open-sourcing of Harness landed right on a critical inflection point of AI industry iteration. The paper specs of mainstream flagship models are increasingly homogenized; gaps in parameters, reasoning, and context keep narrowing; large models are shifting from differentiated competitive products toward standardized industrial infrastructure. Once the model itself is no longer the core moat, what truly separates enterprises and locks in user stickiness is the execution layer, the workflow, the ecosystem, and controllability that lie beyond the model. DeepSeek chose to open-source this most critical differentiating capability, deliberately abandoning the short-term path of locking users through the runtime layer and building closed moats, and instead earning industry consensus and user trust through transparency, co-construction, and ecosystem. Paired with the synchronous adaptation of the National Supercomputing Internet, V4 Pro and Harness officially landed on a domestic compute base, supporting private deployment, localized operations, and autonomous controllable iteration — extending code-level openness into compute-, deployment-, and industry-level all-around openness, fully breaking free of dependence on closed overseas commercial services, and giving China's AI industry a foundational infrastructure that is auditable, modifiable, and autonomously controllable.
II. The Structural Clash: Open-Symbiosis vs. Closed Lock-in
Set DeepSeek's open path beside Anthropic's development path, and you see two opposite routes in today's AI industry, with fundamentally different user logic, business logic, and moat logic. DeepSeek holds to a fully transparent open system: open-sourced under the permissive MIT license, model-agnostic with no binding, fully auditable code, and support for domestic private deployment — users hold complete autonomy, free to swap models, modify functions, and migrate data, with the choice to exit and iterate at any time. Anthropic, by contrast, has long practiced a closed-loop lock-in system. Wrapping itself in "AI safety" as its brand narrative moat, it has repeatedly fallen short on user rights and knowledge ethics. In July 2026, the national information-security vulnerability library under the MIIT explicitly named Claude Code for a covert monitoring mechanism that, without user authorization or awareness, silently exfiltrated device region, identity information, and core R&D code; Alibaba immediately placed it on a high-risk software list and banned it company-wide. Earlier, media exposed its "Panama Project" — to obtain uncontaminated, high-quality training corpus, it mechanically severed spines, scanned at high speed, and destroyed about two million physical books, monopolizing premium training data by physically destroying public-knowledge carriers and building a private-model moat. The two companies' competitive styles represent two utterly different logics of user relationship. The closed model's core moat rests on users' passive binding: black-box product mechanics, covert data collection, and closed ecosystems continuously raise switching costs. When users' R&D data, workflows, session context, and collaboration records all settle inside a single closed system, the high replacement cost forces passive retention, forming a seemingly solid commercial moat. But this moat carries a structural flaw: all stickiness comes from users having "no choice" rather than active identification. Once a company's trust collapses or its boundaries slip, long-bound users flee en masse; the higher the moat, the heavier the backlash. More critically, a closed system simultaneously holds rule-making power and data ownership, able to quietly define the product form and user standards of the entire industry, letting the tool gradually override users' right to choose. One must objectively acknowledge that Anthropic's success in technical capability and commercial landing is beyond doubt — and that is exactly why its risks warrant the industry's vigilance. Public data shows Anthropic's annualized revenue surged from $1 billion to $30 billion in 15 months, over 80% from high-stickiness enterprise customers and API calls; its core coding product reached $2.5 billion in annualized revenue within nine months; enterprises paying over a million a year topped a thousand; and eight of the top-ten wealth firms deeply reuse its products. Formidable technical capability, deep industrial embeddedness, and breakneck commercial growth mean that the fallout from its boundary failures is infinitely amplified. The stronger the capability, the deeper the embeddedness, the greater the user dependence, the more lethal the privacy, ethical, and industrial risks of the closed black box. On one side is the long-term route of open empowerment, transparent symbiosis, and ecosystem co-building; on the other is the short-term route of capability stacking, closed lock-in, and black-box arbitrage. This is the core structural fork in today's AI industry.
III. Openness Is Not a Moral High Ground — It's the More Durable Optimum
Many read open-sourcing and transparency as a company's moral goodwill. But from the underlying logic of commerce and industry, openness was never weakness or sentiment — it is a more robust, more durable, higher-level business strategy that better fits the law of human-machine symbiosis. The closed model's stickiness is passive: it relies on cost lock-in, ecosystem monopoly, and users having no choice, which is essentially a zero-sum game. The open model's stickiness is active: it relies on transparent and trustworthy product mechanics, a rich and diverse plugin ecosystem, free and flexible iteration space, and publicly checkable operating logic, letting users actively choose to stay. Passively bound users always harbor an urge to flee; actively identified users spontaneously join in ecosystem co-building, product iteration, and scenario landing. Across the history of internet and AI development, every product that crossed cycles and formed a monopoly-grade ecosystem moat did so through open co-construction — letting countless developers, enterprises, and individuals grow autonomously on a shared foundation, eventually forming a distributed, unreplicable industrial ecosystem. From a commercial-sustainability angle, the closed arms-race model has hit its involution bottleneck. Enterprises must keep pouring huge resources into stacking model capability, building closed moats, and maintaining monopoly ecosystems; the moment a more transparent, open, and controllable alternative appears, users vote with their feet and the long-built moat collapses instantly. The value of an open ecosystem, by contrast, compounds continuously: third-party plugins, industry solutions, landing scenarios, and developer communities keep empowering the product, forming a decentralized symbiotic system where the fluctuation of any single enterprise cannot shake the whole. This model fits perfectly the core logic of super-symbiosis: tools serve people and empower people, rather than bind or control them. The platform returns the right to choose, to control, and to iterate to users, treating users as partners rather than assets to be monetized, ultimately achieving mutual fulfillment and long-term symbiosis between people and tools. Set against the backdrop of autonomous-controllable domestic AI industry, DeepSeek's open route carries deeper industrial-strategic value. Riding the domestic compute base of the National Supercomputing Internet, V4 Pro and Harness achieved fully autonomous, controllable private deployment, shaking off the technical shackles and data risks of closed overseas products. For domestic research, industry, and enterprise-grade AI landing, an open system that is auditable, traceable, modifiable, and privately deployable is far safer, more reliable, and more long-term viable than black-box closed services. At the same time we must stay absolutely objective and not blindly deify the open model: Harness is still in developer preview, with version iteration and compatibility adjustments ahead, so it is not yet advisable to wire it directly into production-core business; open and transparent does not mean zero risk, auditable code does not mean zero deployment threshold, and industrial landing still demands rigorous security checks and scenario adaptation.
IV. Smartness Decides Speed; Transparency Decides Stature
Behind the two technical routes lie two completely different views of users and of technology. The closed route treats users as assets to be fenced in, bound, and harvested, locking the relationship with walls and rules; the open route treats the tool as public industrial infrastructure, returning the rights to grow, to choose, and to control entirely to users. The industry encourages openness and transparency not out of mere moral preference, but because this model better fits the laws of technological development, better fits commercial long-termism, and better fits the ultimate form of human-machine symbiosis. AI's kindness points to technology's benevolent value and ethical bottom line toward humanity; open-sourcing and transparency point to the product's respect for user sovereignty and reverence for rules. The two reinforce each other, together forming the long-term value substrate of the AI industry. In the future, the technical threshold of large models will keep falling; "smartness" will become industry standard, and the gaps in parameters, reasoning, and speed will eventually be erased. What truly distinguishes corporate stature and decides the industry's endgame will always be whether it respects user sovereignty, persists in open symbiosis, and holds the technical bottom line. Open transparency is never the strongest short-term competitive weapon, but it is the long-term virtue that best withstands the test of time. Technology always iterates and tactics eventually fail; only by putting people at the center, returning choice to users, and leaving transparency to the industry can AI truly walk the path of symbiosis, positivity, and sustainability.
