⚠️ Editorial note: The open source ecosystem in China operates under a distinct institutional framework — characterized by state-led initiatives, intranet-like boundaries, and top-down governance. Readers should be aware that this context differs from the community-driven open source model common in other regions. The term “open source” as used in Chinese media may refer to practices that diverge from the conventional definition.

China Open Source Daily — 2026-08-31

🏛️ Huawei — openJiuwen Receives Four Acceptances at EMNLP 2026 — A First for a 2012-Lab Open-Source Agent Platform at ACL-Family Top-Tier Level

1. 华为开源 / openJiuwen (WeChat) — “openJiuwen多项成果被EMNLP 2026录用” (openJiuwen’s multiple achievements accepted at EMNLP 2026)

The 华为开源 / openJiuwen WeChat official account (mp.weixin.qq.com/s/EgRPMF9XAIisPiGHwx5ukw) published on August 29, 2026 the acceptance of four openJiuwen-team papers at EMNLP 2026 (The 2026 Conference on Empirical Methods in Natural Language Processing, October 24–29 in Budapest, Hungary):

  • 2 papers in Findings (the top ACL-family sub-track below the main conference):
    • “Prompting to Prompt: Meta-Template Learning for Transferable Prompt Optimization” — Zhang Xinyu, Li Deyang, Zhao Minjun, Han Meng, Shi Ruifeng, Dou Zhicheng. Proposes PTP (Prompting to Prompt), a meta-learning-inspired framework that replaces single-prompt optimization with a reusable structured Meta-Template.
    • “Gnosis Deepsearch: Structured Search Space Navigation for Complex QA” — Halil Amut et al. Recasts deep-search as a structured constraint-satisfaction and search-space navigation problem; combined with the LEGO retrieval method, it improves gpt-oss-20b ReAct accuracy from 23.0% → 46.8% → 61.2% on BrowseComp-Plus, and DeepSeek-V4-Flash from 56.2% → 72.2% → 83.4%.
  • 1 paper in Industry Track:
    • “Toollery: Scaling LLM Agents to Thousands of Skills and Tools” — Tian Xiangxi, Guan Ran. A training-free retrieval-based candidate-compression framework that scales agent tool selection to ~79,000 capabilities.
  • 1 system demonstration in Demo Track:
    • “GOD: Govern, Observe, and Direct — A Real-Time Control Room for Agent Societies” — Luo Yige, Guan Ran. A human-in-the-loop multi-agent society simulation tool combining AgentSociety and openJiuwen.

The account closes by identifying openJiuwen as “built jointly by Huawei’s 2012 Laboratory, Huawei Cloud, Terminal, and Computing teams” — an open-source AI agent platform, and points to its flagship multi-agent product WorkSwarm, with entry points at openjiuwen.com, github.com/openJiuwen-ai, and atomgit.com/openJiuwen.

Institutional significance: This is the first documented instance in which a Huawei 2012-Lab–affiliated open-source agent platform has received multiple top-tier ACL-family academic acceptances at the same time — including one Industry Track paper — converting peer-reviewed academic legitimacy into a commercial-product promotion instrument.

From an institutional economics perspective, the openJiuwen EMNLP acceptance event matters for four reasons:

First, it documents the “Industry Track” acceptance as a structurally new institutional-credibility instrument for a Chinese vendor open-source project. Prior Huawei open-source documentation (the August 25 briefing’s CANN move at CCF COSCon; the August 26 briefing’s HCCL contest; the August 30 briefing’s CANN 70,000-RMB task board) operated at the platform-structure / talent-pipeline / contractor-market layers — all internal-Chinese-community credibility instruments. This cycle’s EMNLP acceptances operate at the ACL-family-peer-review layer — the international top-tier-venue credibility instrument that Chinese vendor open-source projects have historically been under-represented in. From an institutional economics standpoint, this domestic-community-credibility / international-peer-review-credibility pairing is a structural finding: Huawei’s openJiuwen has achieved, in a single EMNLP cycle, a dual-credibility posture (AtomGit + GitHub + top ACL-family venue) that prior Huawei open-source initiatives did not document.

Second, the “Industry Track” paper (Toollery) is institutionally the most significant of the four. Findings papers document methodological contributions; Industry Track papers document deployed industrial solutions — and EMNLP’s Industry Track is, per ACL-family norms, the tier at which a peer-reviewed academic paper can explicitly reference an ongoing production system. Toollery’s abstract describes a training-free retrieval-based candidate-compression framework that scales agent tool selection to ~79,000 capabilities and is described as a “training-free engineering path that can be directly embedded in existing pipelines.” From an institutional economics standpoint, this industry-track / production-deployment pairing is a structural finding: the paper is not describing a research artifact but a commercial-product component of Huawei’s own agent stack — the peer-reviewed venue is being used to certify that Huawei’s tool-selection infrastructure is a genuine scaling artifact, not a demonstration-only prototype. This is the first documented instance in this series of a Chinese vendor open-source paper using ACL-family peer review to certify commercial deployment legitimacy.

Third, it documents the 2012-Lab / Huawei Cloud / Terminal / Computing “joint team” authorship as an institutional-form fact. Prior Huawei open-source documentation treated Huawei open-source projects as unitary corporate actors. The openJiuwen WeChat post explicitly identifies the authoring team as a joint formation of four internal Huawei organizations: the 2012 Laboratory (research), Huawei Cloud (services), Terminal (consumer devices), and Computing (chips and infrastructure). From an institutional economics standpoint, this four-organization-joint-team / single-open-source-platform pairing is a structural finding: the openJiuwen platform is not one team’s open-source project but a cross-functional Huawei consortium project — a Williamson L2-level institutional environment in which four internal Huawei cost-centers have been aligned around a shared open-source artifact, and EMNLP acceptance is the external-legitimacy seal on that internal institutional alignment.

Fourth, it documents the “GitHub + AtomGit dual-hosting” posture as a distribution strategy at the academic-credibility layer. The openJiuwen post cites github.com/openJiuwen-ai and atomgit.com/openJiuwen in parallel — the same dual-hosting posture that prior briefings have documented on the CANN / HCCL codebases. The EMNLP paper’s reference to openJiuwen implicitly certifies the dual-hosting distribution model to the international peer-review community. From an institutional economics standpoint, this dual-hosting / ACL-credibility pairing is a structural finding: the international academic community is being asked to treat the AtomGit copy as legitimate, not just the GitHub copy — a dual-hosting legitimacy transfer event that prior briefings did not document.

Sources:

🔍 Jiang Tao (硅基时间/Silicon Time) — Part II of “Chinese Large Models Are World-Class, Why Is the Developer Ecosystem Still in Others’ Hands?” — The First Quantified Derivative-Mass Gap for the Default-Path Thesis

2. 硅基时间 / Silicon Time (Jiang Tao, WeChat) — “中国大模型已经世界一流，为什么开发者生态还在别人手里？（下篇）” (Chinese large models are world-class — so why is the developer ecosystem still in others’ hands? — Part II)

The 硅基时间 / Silicon Time WeChat official account (mp.weixin.qq.com/s/Z6fbNJr4qTw8MGJoKwXgUg) published on August 30, 2026 Part II of Jiang Tao’s (蒋涛) two-part essay on the developer-infrastructure control-point problem in Chinese AI. Part I (referenced but not reproduced here) addressed Nvidia’s $12.9 billion purchase of Hugging Face and argued that what Nvidia bought was “model distribution + developer relations + compute routing” control points, not just a model repository. Part II addresses the downstream question: given that China has world-class models and rapidly building chips, why does the developer ecosystem’s default path still run through U.S. platforms and tools?

The five core claims of Part II (from Jiang Tao’s own summary):

  • (1) China has world-class large models and rapidly building chips — what it lacks is the middle layer of toolchain, developer-distribution platform, and long-term capital willingness to fund that layer.
  • (2) Adaptation capability is no longer weak — the real gap is the ecosystem multiplier (生态乘数): whether one model release can automatically grow thousands of quantized / fine-tuned / on-device / chip-adapted derivatives, and whether those derivatives can be re-read, registered, and redistributed by the default toolchain.
  • (3) The middle layer has long had no name — and its value largely overflows out of the platform: it flows to chip-load growth, model-ecosystem growth, tool-company growth, and investment returns — so it cannot be priced on platform revenue alone; four account books must be kept simultaneously.
  • (4) Copying Hugging Face has no path — China’s three cards are open models + heterogeneous compute + Physical AI industry scenarios + the global-neutral-layer vacancy; the two control points to win are horizontal “Run Anywhere” (where AI runs) and vertical “Agent Runtime” (how AI works).
  • (5) The first step is not to build another big platform — first build the public layer of “model × chip” compatibility/performance/cost so developers can point-and-run on any chip. Then let the platform grow out of real usage.

The empirical anchor Jiang Tao provides — and which the Part I briefings did not quantify:

  • Qwen on ModelScope: ~34,000 derivatives. Qwen on Hugging Face: 151,000 derivatives.
  • Of the 28,000 GGUF-conversion versions of Qwen on Hugging Face, Qwen itself published only 54 — the rest are produced by the international developer-toolchain community.
  • Qwen was published on ModelScope 24–48 hours before Hugging Face, with free inference and free GPU-time subsidies on ModelScope — and the derivative mass still went overwhelmingly to Hugging Face.
  • AtomGit claims 12 million developers; ModelScope claims 10 million+ users and 170,000 open models. Both are structurally tied to a single chip vendor or a single cloud vendor, which Jiang Tao argues is sufficient for distribution layer but structurally insufficient for compute-routing / chip-adaptation / settlement layers.
  • Hugging Face’s user-geography is not primarily U.S. users — third-party traffic data shows most Hugging Face users come from global non-U.S. regions, and Chinese models account for 41% of Hugging Face’s total downloads. The moat is not the U.S. market; it is the global developer network.
  • The Qwen-team day-zero direct grant of access to Unsloth is cited as an admission that Chinese teams already know: to enter global quantization, fine-tuning, and local-runtime ecosystems on day one, they must directly plug into already-existing international toolchains — because llama.cpp, Ollama, LM Studio, Unsloth, and mlx-community default-read Hugging Face.
  • The mechanism by which Hugging Face became the default platform is technical, not market-share-based: the 2018 open-sourcing of the transformers PyTorch codebase hard-coded two commands — one to download from Hugging Face, one to upload back to Hugging Face — and global developers write those two lines daily.

Jiang Tao’s concluding framing: “The gap is not in one adaptation’s capability, but in the industrial capability to replicate one adaptation into thousands of public assets and let global tools default-read them.” And: “Model industry influence ≈ model capability × developer-ecosystem multiplier. China has demonstrated the first factor; what is missing is the second.”

Institutional significance: This is the first quantified empirical anchor for the “default-path / derivative-mass” thesis that prior Silicon Time briefings referenced but did not numerically ground.

From an institutional economics perspective, the Silicon Time Part II publication matters for five reasons:

First, it documents the derivative-mass gap (34,000 vs. 151,000) as the first publicly-published quantified anchor for the “ecosystem multiplier” thesis — a Williamson L3 reputational-incentive-layer observation quantified for the first time. Prior Silicon Time briefings (August 20, 21, 23, 24, 25) documented Silicon Time’s commentary on Jeff Dean’s departure, Hugging Face’s global dominance narrative, and the general “developer ecosystem control point” thesis — but did not attach specific numerical anchors to the claim. This cycle’s Part II attaches three numerical anchors: the 34,000 / 151,000 Qwen-derivative gap, the 28,000 GGUF-with-only-54-Qwen-authored breakdown, and the 24–48 hour ModelScope-first-release-window. From an institutional economics standpoint, this thesis / numerical-anchor pairing is a structural finding: the “ecosystem multiplier” argument, previously a narrative, is now empirically anchored — and the anchor sits materially in favor of the thesis (34,000 / 151,000 ≈ 22.5%, i.e. ModelScope captures only ~22.5% of the Qwen derivative ecosystem that Hugging Face does despite holding the first-release window and subsidies).

Second, it documents the “transformers library hard-coded two-line command” mechanism as an institutional-form fact about how defaults form — an observation that resonates with the Williamson L1 social-embeddedness layer. Jiang Tao’s core argument for why the middle layer cannot be built by a copy of Hugging Face is that Hugging Face’s dominance is not market-share-based but technical-embeddedness-based: the 2018 open-sourcing of transformers hard-coded two commands (one download, one upload) into a library that global developers write to daily, and after years of use the two commands became the default path. From an institutional economics standpoint, this hard-coded-command / L1 social-embeddedness pairing is a structural finding: the “default path” is not a market-share result but a technical-inertia result — a Hayekian spontaneous-order observation in which no one chose Hugging Face as the default, but the sum of millions of small developer decisions to write the same two lines of code produced a self-reinforcing default. The institutional implication — which Jiang Tao draws explicitly — is that a copy of Hugging Face cannot reverse a Hayekian spontaneous order; only a different technical-embedding mechanism can.

Third, it documents the “day-zero direct grant of access to Unsloth” as an institutional-form admission by Chinese model labs that the international toolchain is the day-zero path. Jiang Tao explicitly notes: “The Qwen team proactively granted day-zero access to Unsloth.” This is an institutional-form admission — not a market decision but a strategic decision by a Chinese model lab to bypass its own domestic toolchain layer in order to enter the international toolchain layer on day zero. From an institutional economics standpoint, this domestic-platform-first-release / international-toolchain-day-zero pairing is a structural finding: Chinese model labs are structurally aware that their own platform layer is not yet a sufficient distribution and differentiation vehicle, and they are willing to bypass their own domestic platforms to enter the international ecosystem immediately. The domestic platforms (ModelScope, AtomGit) are simultaneously (a) the venues where the release happens and (b) not the venues where the ecosystem multiplier operates.

Fourth, it documents the “AtomGit + ModelScope single-vendor-tie” argument as an institutional-form critique of China’s two main domestic platforms. Jiang Tao explicitly argues that AtomGit and ModelScope are structurally tied to a single chip or cloud vendor each, and that this single-vendor tie is sufficient for the distribution layer but structurally insufficient for the compute-routing / chip-adaptation / settlement layers. AtomGit is described as the official platform of the OpenAtom Foundation with 12 million developers (the OpenAtom Foundation being one of the four open-source foundations tracked by prior briefings); ModelScope is described as an Alibaba-Cloud-owned model community with 10 million+ users and 170,000 open models. From an institutional economics standpoint, this foundation-governance-platform / cloud-vendor-platform pairing is a structural finding: China’s two main domestic model platforms operate under two different Williamson institutional forms — AtomGit under a foundation-governance form (OpenAtom Foundation), ModelScope under a single-vendor-cloud form (Alibaba Cloud) — and Jiang Tao’s argument is that neither form is structurally sufficient to own the compute-routing layer, because compute-routing requires vendor-neutrality that neither foundation nor single-vendor cloud can provide.

Fifth, it documents the “four-account-book” argument as a structurally new institutional-economics framing for why the middle layer has long been under-funded. Jiang Tao’s third claim is that the middle layer has long had no name — and that its value largely overflows out of the platform, flowing to (a) chip-load growth, (b) model-ecosystem growth, (c) tool-company growth, and (d) investment returns — so the platform layer cannot be priced on platform revenue alone, and four account books must be kept simultaneously. From an institutional economics standpoint, this platform-revenue / four-account-book pairing is a structural finding: it identifies the institutional cause of the middle-layer funding gap as a Coasian transaction-cost / property-rights-mismatch problem — the platform owner captures the direct platform revenue but the externalities (chip-load, model-ecosystem, tool-company, investment) accrue to non-platform actors. This is a Coase-North framing that prior Silicon Time briefings did not articulate.

Sources:

🏛️ Xie Lan (清华AIRI院长, 国务院参事) — “答好时代之问，引领全球人工智能治理” in 《中国社会科学报》 — Top-Tier Academic Authority Formally Adopts “Great Divergence” in AI Governance Discourse

3. 清华大学人工智能国际治理研究院 (Tsinghua AIRI) / 中国社会科学报 — “薛澜 王净宇：答好时代之问，引领全球人工智能治理” (Xie Lan, Wang Jingyu: Answering the question of the era, leading global AI governance)

The Tsinghua AIRI (清华大学人工智能国际治理研究院) WeChat account republished on August 30, 2026 an article in 《中国社会科学报》 (China Social Sciences News) — one of the most authoritative official social-science media in China — co-authored by Xie Lan (薛澜) (State Council Advisory Group member, Dean of Tsinghua’s Shapiro Institute for Global Advisors, Director of Tsinghua AIRI, Director of the China Science and Technology Policy Research Center) and Wang Jingyu (王净宇) (Assistant Researcher at Tsinghua AIRI).

Xie Lan’s authorship is institutionally significant: he is the senior-most academic voice in the Chinese AI-governance discourse — a State Council Advisory Group member, and Director of the academic institute that is effectively the official Chinese voice on global AI governance. Any institutional term he uses becomes a policy-discourse reference point for downstream research and policy.

Three institutional signals documented in the article:

Signal 1 — The “Great Divergence” term enters the top-level AI governance discourse. Xie Lan uses the term “大分流” (Great Divergence) directly: “长此以往，资源、能力与文化等多重鸿沟相互强化、不断扩大，新一轮国家间的**‘大分流’**现象可能成为现实。” (Over time, the multiple gaps in resources, capabilities, and culture mutually reinforce and continuously widen, and a new round of the international “Great Divergence” phenomenon may become reality.) “Great Divergence” is the term popularized by Kenneth Pomeranz’s The Great Divergence (2000) — the history-of-economics concept explaining the West’s 18th-century economic take-off from China and Europe. Its appearance in a 《中国社会科学报》 article by Tsinghua AIRI’s dean is a formal-adoption event: the Great Divergence 2.0 analytical framework — the framework this daily briefing has used as its interpretive lens since inception — has been formally adopted by China’s senior-most AI-governance academic voice as a policy discourse term.

Signal 2 — Open source is explicitly positioned as a diplomatic instrument and cultural-sovereignty path. Xie Lan makes two explicit judgments about open source’s role in Chinese statecraft:

Original quoteInstitutional meaning
“以开源模式和开放平台降低技术使用门槛，帮助全球南方国家加强能力建设” (Use open-source models and open platforms to lower the technical barrier to use, helping global South countries strengthen capacity building)Open source is a diplomatic instrument
“以开源模式开放高质量基础模型，帮助各国在自主可控基础上发展符合本土文化与伦理的应用” (Open high-quality foundation models through open source, helping various countries develop applications that meet local culture and ethics on the basis of autonomy and control)Open source is a cultural-sovereignty implementation path
“个别国家借机推广绑定本国技术栈的做法名为赋权，实为依附” (Some countries use this opportunity to promote practices that bind to their own tech stacks — called ’empowerment’ but actually ‘dependency’)U.S. AI exports are reframed as “tech-stack lock-in”; Chinese open source is reframed as genuine autonomy

From an institutional economics standpoint, Xie Lan’s framing is the most explicit yet from a Chinese senior academic authority on open source as a statecraft instrument — not as a technical practice, not as a commercial model, but as a diplomatic and cultural-sovereignty tool. The framing is institutionally coherent with the August 12 memory note on “open source is a club good, not a public good” — Xie Lan’s article treats open-source-model distribution as a statecraft tool that confers diplomatic standing and cultural sovereignty, not as a technical contribution to global commons.

Signal 3 — The “world AI governance” framework is presented as a “China Plan” (中国方案) with five pillars. Xie Lan articulates a “China Plan” with the following six pillars:

  • (1) Build a new-type multilateral cooperation platform (the World AI Cooperation Organization, established at WAIC in Shanghai in July 2026).
  • (2) Play the great-power coordination role, particularly on early-harvest items with the U.S.
  • (3) Build a tiered-and-coordinated governance landscape (UN for capacity building, ISO/research institutions for technical standards, great-power bilaterals for high-sensitivity issues).
  • (4) Use inclusive cooperation to bridge the digital intelligence gap.
  • (5) Use open-source models + open platforms to promote multi-cultural inclusive development (the pillar that contains the open-source-as-statecraft statements above).
  • (6) Jointly build the safety governance bottom line.

The structural institutional fact here is that “China Plan” is now an articulated six-pillar framework in 《中国社会科学报》 — the top-tier official media — attributed to a State Council Advisory Group member. This is a formal discourse-institutionalization event for Chinese AI governance policy.

Institutional significance: This is the first documented instance in which China’s senior-most AI-governance academic voice, in China’s most authoritative social-science official media, has simultaneously (a) adopted the “Great Divergence” term in AI-governance discourse, (b) positioned open source as an explicit instrument of diplomatic statecraft and cultural sovereignty, and (c) articulated a “China Plan” for global AI governance as a six-pillar framework.

From an institutional economics perspective, the Xie Lan / China Social Sciences News publication matters for five reasons:

First, the “Great Divergence” adoption is a formal discourse-legitimation event for the Great Divergence 2.0 framework. The Great Divergence 2.0 analytical lens — the framework this daily briefing has used since inception to distinguish FLOSS (真开源) from State-Chartered Codebase / Intranet Shared Source / Cyber-Estate (伪开源) — has been introduced into the top-tier Chinese AI-governance academic discourse by its senior-most voice. From an institutional economics standpoint, this analytical-lens / discourse-adoption pairing is a structural finding: the Great Divergence 2.0 framework, previously an external analytical lens, is now a term used by the Chinese policy-academic establishment itself — and the adoption signals that Chinese policy-academic discourse has itself recognized the “new-round Great Divergence” as a real institutional-risk scenario in AI.

Second, the “open source as diplomatic instrument” positioning is the most explicit yet from a Chinese senior academic authority — and it is institutionally coherent with the FLOSS / State-Chartered Codebase distinction. Xie Lan’s article positions open source as a diplomatic instrument that “helps global South countries strengthen capacity building” and “helps various countries develop applications that meet local culture and ethics on the basis of autonomy and control.” The counter-framing — that U.S. AI exports are “called empowerment but actually dependency” — is a direct institutional claim about the difference between (a) Chinese open source (framed as autonomous capability-building) and (b) U.S. proprietary AI exports (framed as tech-stack lock-in). From an institutional economics standpoint, this Chinese-FLOSS / US-proprietary-lock-in framing is a structural finding: Xie Lan’s article positions the Great Divergence 2.0 distinction — FLOSS as a capability-building instrument vs. proprietary stack as a lock-in instrument — as a policy-relevant institutional choice. The framework this daily briefing has used as its interpretive lens is now, in Xie Lan’s article, an explicit Chinese policy framing.

Third, the “China Plan” six-pillar framework is a formal discourse-institutionalization event for Chinese AI governance policy. The “China Plan” (中国方案) has long been an implicit Chinese policy framing. Xie Lan’s article, in 《中国社会科学报》, articulates the “China Plan” as an explicit six-pillar framework, with the fifth pillar (open source + open platforms for cultural diversity) as the pillar that contains the open-source-as-statecraft statements. From an institutional economics standpoint, this implicit-framework / explicit-framework pairing is a structural finding: the “China Plan” for global AI governance has moved from implicit to explicit, and the explicit form is now documented in China’s most authoritative social-science official media, attributed to a State Council Advisory Group member.

Fourth, the “world AI governance from consensus to institution-building” thesis is a structurally new institutional-economics framing for Chinese policy. Xie Lan’s opening claim is that “global AI governance has moved from a consensus-formation phase to an institution-building phase, and the pace of governance is markedly accelerating.” The article then documents a specific institutional-timeline: February 2026 (89 countries + 2 international organizations sign the India AI Impact Summit declaration), May 2026 (China–U.S. summit agrees to AI intergovernmental dialogue), July 6, 2026 (first UN Global Dialogue on AI Governance in Geneva), July 17, 2026 (Xi Jinping’s main speech at WAIC and establishment of the World AI Cooperation Organization in Shanghai). From an institutional economics standpoint, this consensus-phase / institution-building-phase transition thesis is a structural finding: the Chinese policy-academic establishment is now formally recognizing a phase transition in global AI governance, and positioning Chinese institutional initiatives (WAIC, World AI Cooperation Organization) as institutional contributions to that phase transition.

Fifth, the article’s citation of third-party institutional endorsements documents a structurally new external-legitimacy posture for the “China Plan.” Xie Lan’s article cites: Kazakhstan President Tokayev’s evaluation that China plays a leading role in AI open-source innovation, open platform construction, and international tech cooperation; UN Deputy Secretary-General Gill’s statement that China is a “role model for global AI capacity building”; MIT’s Max Tegmark’s evaluation that China’s equal-priority of safety and development “sets an important example for governments.” From an institutional economics standpoint, this China-Plan / third-party-endorsement pairing is a structural finding: the “China Plan” is not being presented as an isolated Chinese claim but as a plan that has received specific third-party institutional endorsements — from a Central Asian head of state, a UN deputy secretary-general, and a Nobel-affiliated MIT researcher — and the endorsements are cited as institutional legitimacy for the “China Plan” framework.

Sources:

🔍 Commentary

Three institutional moves in one cycle — one shared unsolved object: “the middle layer.”

This cycle’s three stories document, on three different institutional surfaces, the same unsolved institutional object: the middle layer between production and distribution.

  • openJiuwen at EMNLP documents the middle layer at the academic-credibility layer — a Huawei 2012-Lab open-source agent platform using ACL-family peer review to certify commercial deployment legitimacy.
  • Jiang Tao’s Silicon Time Part II documents the middle layer at the developer-ecosystem-control-point layer — quantifying the 34,000 / 151,000 Qwen-derivative-mass gap, and diagnosing the Coasian transaction-cost / property-rights-mismatch cause as the “four-account-book” problem.
  • Xie Lan in 中国社会科学报 documents the middle layer at the policy-discourse layer — the top-tier Chinese AI-governance academic voice adopting the “Great Divergence” term, positioning open source as an explicit instrument of diplomatic statecraft and cultural sovereignty, and articulating a “China Plan” for global AI governance.

The three events together document that the middle layer — between model/chip production and global developer distribution — is the shared unsolved institutional object across all three institutional surfaces: openJiuwen is trying to legitimize its industry-track deployment through international peer review; Jiang Tao is diagnosing the middle-layer funding gap as a property-rights mismatch; Xie Lan is framing open-source-model distribution as a policy instrument to bridge the middle-layer gap for global South countries.

Three institutional moves on three Williamson layers.

Reading the three stories on Williamson’s L1→L4 axis:

  • openJiuwen EMNLP operates at L3 (governance mechanism) — the peer-review institution is a governance mechanism that certifies commercial deployment legitimacy.
  • Jiang Tao’s Silicon Time operates at L2 (institutional environment) — the Coasian transaction-cost / property-rights-mismatch diagnosis is an institutional-environment observation about why the middle layer is under-funded.
  • Xie Lan in 中国社会科学报 operates at L1 (social embeddedness) — the “Great Divergence” term’s adoption, the “China Plan” articulation, and the open-source-as-statecraft positioning are all L1 social-embeddedness events: discourse-level movements that shape how the Chinese policy-academic establishment understands the middle-layer problem.

One perspective, not a verdict.

All three stories — openJiuwen’s EMNLP acceptance, Jiang Tao’s Silicon Time Part II, and Xie Lan’s 中国社会科学报 article — are best read as observations of institutional movement in progress, not as verdicts on institutional direction. openJiuwen’s EMNLP acceptance does not guarantee industrial success; Jiang Tao’s derivative-mass-gap diagnosis may be structurally correct but does not identify a Chinese institutional solution; Xie Lan’s “China Plan” articulation does not resolve the international-governance coordination problem it identifies. What this cycle’s briefing documents is that all three moves have crossed a first-documented threshold — openJiuwen’s is the first 2012-Lab open-source agent platform at ACL-family top-tier; Jiang Tao’s is the first numerically anchored derivative-mass-gap thesis; Xie Lan’s is the first top-tier Chinese AI-governance academic voice to formally adopt the “Great Divergence” term — and that all three thresholds sit on the same unsolved institutional object: the middle layer.

Editorial note on perspective: This briefing presents one institutional-economics reading of Chinese open-source developments, not a verdict. The “institution” in these stories — the vendor, the developer community, the policy-academic establishment — is treated as an object of observation, not a target of critique. The Great Divergence 2.0 framework (FLOSS vs. State-Chartered Codebase vs. Intranet Shared Source vs. Cyber-Estate) is a lens, not a universal answer. One perspective, not a verdict.