Intelligence notes

◉Dashboard↗

The Open Frontier — China's AI Models as Instruments of State Diffusion

19 min read

The Open Frontier — China’s AI Models as Instruments of State Diffusion

Strategic Intelligence Assessment | intelligencenotes.com

Bottom Line Up Front

As of mid-2026, China’s frontier-model ecosystem has done two things at once: it has converged to within single digits of the Western closed frontier on composite capability benchmarks, and it has seized the global open-weight layer outright. (Fact / High.)

What matters is not the narrowing gap but the inversion it reveals. Where Western leaders gate capability behind vetted, opaque access tiers — the two-tier model in Fable 5 and the Two-Tier Frontier, where the safeguarded public model and its brakes-off twin are the same capability under two governance configurations — China’s leading labs do the opposite. They diffuse capability through permissively licensed open weights, given to anyone who will build on them. (Assessment / High.)

Open diffusion is not a weaker move than parity-chasing but a different one, converting the model layer into an instrument of standard-setting, developer dependency, and — because guardrails are trained into the weights, not bolted onto a server — exported value-alignment. The two regimes are mirror images of one logic: both the West and China have made the frontier model an object of access governance. The West gates the unsafeguarded tier and withholds it; China diffuses the safeguarded-to-state-values tier and propagates it. (Assessment / High.) That symmetry, and its embedded cognitive-warfare exposure, is what this assessment tracks.

1. The Convergence — Capability

China’s frontier landscape in mid-2026 is a dense field of model families, several leading the open-weight world. Figures below are lab-self-reported unless aggregator-tied:

LabLatest model (mid-2026)Scale (total / active)LicensePosture
DeepSeekV4-Pro / V4-Flash (Apr 2026)1.6T / 49B · 284B / 13BMIT (per secondary; license text not located — Gap)Open-weight leader
Alibaba QwenQwen 3.7 Max (closed) / Qwen3 (open)—Apache-2.0 (open tier)Closed #5 global + dominant open base
Zhipu / Z.aiGLM-5744B / 40BMITSWE-bench Verified 77.8% (self-reported)
MoonshotKimi K2.x1T / 32BModified MIT#1 open-weight on Artificial Analysis Index
BaiduERNIE 4.5 (open) / 5.x (closed)—Open tier permissiveSplit open/closed
ByteDanceDoubao Seed-OSS-36B (open) / Seed 2.0 (closed)36B (open)Open tierSplit open/closed

Three further labs round out the field: MiniMax (open, efficiency-focused), StepFun (open-weight entrant), and Tencent Hunyuan, whose Hy3 ships under a non-OSI-permissive custom community license — “open” here spans from MIT-grade freedom to use-restricted release. 01.AI exited frontier pretraining for the application layer in 2025.

Two findings stand out. First, the open/closed split — Alibaba, Baidu, and ByteDance each ship a frontier-class closed model and a permissive open one — is a deliberate two-track posture, not a timing accident. (Assessment / High.) Second, the 01.AI pivot is a consolidation signal — the field concentrating around labs with state-scale capital and compute. (Assessment / Medium.)

On the Artificial Analysis Intelligence Index (v4, June 2026) — a third-party aggregator weightier than any self-report — the Western frontier sits around 60–65. The top Chinese closed model, Qwen 3.7 Max, scores 56.6 (#5 globally). The top Chinese open model, Kimi K2.6, scores 54, ahead of DeepSeek V4-Pro at 52 and V4-Flash at 47. (Fact / High.) The best Chinese open models therefore sit 6–11 points below the absolute frontier — a single-digit gap on a 100-point scale, closed faster than most 2025 forecasts projected — and hold roughly four of the top five open-weight slots. (Fact / High.)

The containment camp reads it differently: NIST CAISI assessed DeepSeek V4-Pro in May 2026 as roughly eight months behind the US frontier — read as evidence the controls work. (Fact / Medium.) Both hold — a single-digit composite gap and an eight-month lag at the leading edge.

One constant runs across every Chinese lab: all ship Mixture-of-Experts architectures with low activation ratios — roughly 3–5% of parameters active per token. (Fact / High.) The standard reading — the signature of training under compute scarcity, making each FLOP carry more model — must be carried as Assessment, not Fact: SemiAnalysis cautions that low-activation MoE and the broader efficiency gains are industry-wide algorithmic progress, visible in Western labs too, not a uniquely Chinese response to CHIPS Act-era controls. Attributing the architecture to export controls is an inference, not a measurement. (Assessment / Medium.)

2. The Chokepoint and the Sovereignty Response — Compute

The constraint shaping the architecture is Nvidia silicon and the policy gating it. The export-control timeline (Fact / High) is anchored to Nvidia SEC filings and BIS guidance:

  • Oct 2022 — A100/H100 banned to China.
  • Oct 2023 — the cut-down A800/H800 (built to clear the 2022 threshold) banned in turn.
  • Apr 2025 — H20 (the China-market Hopper part) banned, triggering a $4.5B Nvidia inventory charge; re-licensed Aug 2025 under a ~15% revenue-share, but China-side guidance discouraged its use and only ~$50M of demand materialized.
  • Jan 2026 — H200 moved to case-by-case licensing; 400k+ units approved, none converted. Blackwell denied outright.
  • May 13, 2025 — Huawei Ascend 910B/C/D globally prohibited under BIS General Prohibition 10.

The chokepoint is real, durable, and the one frontier input China cannot route around at home. The domestic response is brute force: Huawei’s Ascend 910C and CloudMatrix 384 supernode, claimed at 1.7× a GB200 NVL72 rack’s throughput at 3.9× the power draw — self-reported, no independent MLPerf corroboration. (Fact / Medium.)

Silicon remains the binding limit. SMIC is stuck at a 7nm-class node via DUV multi-patterning, no EUV lithography — confirmed by TechInsights teardowns. (Fact / High.) Ascend is not yet production-stable for frontier pre-training: gradient-synchronization and operator failures confine it to post-training — DeepSeek post-trained V4-Pro on roughly 1,000 Ascend 910C but pre-trained elsewhere. (Fact / Medium.) The load-bearing distinction: domestic compute can finish a frontier model but cannot yet reliably build one.

Hence the workarounds. Offshore training on Nvidia chips in Singapore and Malaysia is confirmed for Alibaba and ByteDance; domestically-training DeepSeek is the exception proving the pattern. (Fact / High.) A DOJ-confirmed $160M smuggling ring drew a guilty plea in Oct 2025. (Fact / High.) And the demand-side control that would have governed chip flows — the Biden “AI Diffusion Rule” (Jan 2025) — was rescinded by the Trump administration on May 13, 2025, before it ever took effect. (Fact / High.)

2.5 The Embedded Guardrail — Governance and Censorship

This section is the analytic hinge: the compute story explains the architecture; the guardrail story converts an open Chinese model into an instrument of cognitive warfare.

The regulatory floor. China’s CAC Interim Measures for Generative AI Services (effective 2023-08-15), Article 4, requires providers “uphold socialist core values” (坚持社会主义核心价值观) and not generate content “inciting subversion of state power.” ([primary, state, ZH].) The registry is accelerating — 748 registered generative-AI services by end-2025, against 64 in 2023. (Fact / High.) A CN/EN framing delta is significant: CAC’s English communications frame these obligations as technocratic safety — “truth, accuracy, objectivity” — while the Chinese text is explicit about protecting state power; the political-censorship dimension in the ZH original is omitted in the EN relay — an information-confrontation move presenting the same regulation as consumer protection abroad, regime protection at home. (Fact / High.)

The censorship is weight-level, not a server-side filter — the decisive finding, since it determines whether self-hosting escapes the guardrail. It does not. R1dacted (arXiv 2505.12625) found 68.75% of tested sensitive categories 100% censored even in self-hosted deployments; Pan & Xu (PNAS Nexus, Feb 2026) measured Chinese-language political-question refusal rates of BaiChuan 60.2% / DeepSeek ~36% / Ernie 32% / ChatGLM 10%, against 0–2.8% for non-Chinese models. (Fact / High.) Sensitive content often surfaces in the chain-of-thought, then is suppressed in the final answer — the model “knows but is trained not to say.” (Fact / High.) Being in the weights, it ships with the download and survives any local hosting.

Two qualifiers keep it honest. First, reversible: Perplexity’s “R1 1776” de-censored DeepSeek-R1 with under 1% benchmark loss — a thin SFT layer, cheap to add and strip. (Fact / High.) Second, it is language-dependent: R1dacted measured 99.57% censorship on Chinese prompts versus 61.16% on Farsi; Pan & Xu found language, not server location, predicts sanitized output. (Fact / High.) The guardrail is therefore differential audience management — more open to international/English audiences than the domestic/Chinese one; all three properties matter for §5.

3. Open Weights as Statecraft

The defining move is to give the capability away — read through the strategy spine and diffusion metrics together.

The strategy spine is documented. The 2017 New Generation AI Plan (AIDP) set the goal of China as “the world’s primary AI innovation center” by 2030. ([primary, state, ZH]; translation via DigiChina.) Military-civil fusion (军民融合) is written into the AIDP and chaired by Xi; “national team” (国家队) AI Open Innovation Platform designations channel state direction to specific firms; the 2025 “AI+” initiative (State Council Doc [2025] No. 11) operationalizes diffusion across sectors; and the 15th Five-Year Plan (Mar 2026) mentions AI roughly 52 times versus ~6–11 in the 14th. (Fact / High.) State capital now reaches even DeepSeek — a 2026 round drew state funds, Big Fund III, and Tencent; Liang Wenfeng met Xi in Feb 2025. Zhipu/Z.ai is the clearest case: Shanghai state-funded, on the US Entity List. (Fact / High.)

The diffusion metrics are the study’s most striking evidence — Fact / High:

  • Chinese models were 41% of all Hugging Face downloads in 2025.
  • Chinese developer share overtook the US (17.1% vs 15.86%).
  • Qwen reached the high hundreds of millions of cumulative downloads (approaching 1B), overtook Meta’s Llama in Oct 2025, and anchors ~113k–200k+ derivative models — more than Google and Meta combined — with roughly 69% of Hugging Face fine-tunes Qwen-derived. (The ~700M figure is relayed via Xinhua [state-aligned] — labeled accordingly.)
  • On OpenRouter, Chinese-model token share averaged ~13% weekly across a 100-trillion-token year, peaking near 30%. A circulating Feb-2026 “61%” figure is a single-snapshot number — not steady state. (Medium.)

Global South diffusion is documented at case level. AI Singapore switched its SEA-LION sovereign base from Llama to Alibaba Qwen3 in Nov 2025, citing ~8.4% better Southeast-Asian-language performance; Qwen anchors roughly 2,800 derivative apps, including in Brazil and Uganda. At the World AI Conference (Jul 2025) China proposed a Global AI Governance Action Plan and a World AI Cooperation Organization — diffusion as tech diplomacy, dovetailing with the Digital Silk Road and Belt and Road Initiative. (Fact / High.)

This is the “Android of AI” thesis: a free, capable base layer drives developer lock-in, then standards, dependency, and revenue. The logic is well-supported (MIT Technology Review, Lambert, Xu); the coinage is thinly attributed and carried as Assessment, not a sourced phrase.

The intent/effect distinction matters most, since it is where commentary overclaims. The ecosystem-dominance effect is documented Fact: Chinese open models dominate the global open-weight layer. The deliberate-state-direction intent is Western-inferred: no located PRC primary document frames open-weighting itself as state strategy, and at least one Chinese Academy of Sciences researcher’s ZH text frames it instrumentally — a workaround to the chip embargo, not a grand design. On the evidence, open diffusion is an effect that serves the state and is increasingly state-funded — not demonstrably a centrally coordinated plan. (Assessment / Medium.) Holding that line is the difference between analysis and alarm.

4. The Cognitive-Warfare Vector

Here the convergences meet — why the open frontier is a cognitive-warfare question, not a benchmark race. Given §2.5, the mechanism is straightforward: when Global South developers build on Chinese open models carrying PRC-aligned, weight-level guardrails, the embedded framing propagates downstream into every derivative — “censorship-by-design,” exported as “censorship-as-a-service.” The 2,800 Qwen-derived apps and the SEA-LION pivot are that propagation in motion. (Assessment / High on the mechanism.)

The alignment goes beyond refusal. The CSIS Futures Lab CFPD benchmark found DeepSeek-V3 recommends more aggressive foreign policy when the actor is a Western democracy than China or Russia — alignment as analytical posture, not just topic avoidance. (Fact / High.) NIST CAISI and CEPA found DeepSeek echoed roughly four times as many CCP narratives as US reference models, with framing artifacts across English, Chinese, Russian and other languages (Fact / High); the China Media Project’s unreplicated “thought-token forcing” work extracted apparent Qwen3 directives (“Focus on China’s achievements”) (Medium). This is the Three Warfares logic — public-opinion, psychological, legal — at the model layer, propagating through the information environment one derivative app at a time.

The load-bearing gap: the causal chain from model adoption to measurable population-level cognitive effect is unproven. Every link is documented — the guardrail is real, weight-level, default-on, propagating into Global South deployments — but no peer-reviewed end-user behavioral study yet demonstrates a population-scale effect on what people believe or decide. (Gap.) The mechanism is documented, the effect inferred — real in mechanism, unproven in outcome; hold both.

5. The Western Response and the Strategic Debate

The Western frame was set by the DeepSeek “moment”. R1’s release on 2025-01-20 triggered a 17% single-day Nvidia drop ($590B) on 2025-01-27 and “Sputnik moment” framing. (Fact / High.) One year on, recalibration is near-complete: Nvidia recovered, and Big Tech announced roughly $475B in 2026 capex — read most credibly as a demand signal, not a US defeat. (Fact / High.) DeepSeek’s self-reported $5.576M V3 training cost is contested — CSIS and SemiAnalysis put the true all-in cost far higher. (Fact / High.)

The diffusion-vs-containment debate has two coherent schools. Containment — Amodei and Anthropic’s “2028 AI Leadership” — argues for tighter controls to preserve a US lead. The security merits are genuine; so is the structural self-interest that must be flagged — a closed-model lab arguing to restrict its open competitors makes a security case that also defends its business model. Naming it does not dismiss the argument. Diffusion/compete — John Villasenor (Brookings) — argues controls may accelerate Chinese efficiency by forcing §1’s algorithmic innovation, and that open-model bans are in practice unenforceable. (Fact / High on the positions.)

Two precedents discipline the debate. The ChatBIT case — PLA-affiliated researchers fine-tuned Meta’s Llama-13B on military data in 2024 — shows US open weights are equally uncontrollable after release: whatever China gains by diffusing Qwen, the US conceded by diffusing Llama. (Fact / High.) The 2026 distillation escalation sharpened it: Anthropic disclosed to Congress (Feb 2026) that DeepSeek, Moonshot and MiniMax ran roughly 24,000 fraudulent accounts for ~16M extraction exchanges, which OSTP called “industrial-scale.” Whether these are state-directed or commercial competitive intelligence is unresolved in open source. (Fact / High on the disclosure; attribution Gap.) Allied responses have been data-protection-led, not capability-led: Italy’s Garante issued the first ban (Jan 2025), South Korea suspended DeepSeek, and ~eight US states plus federal device bans followed. (Fact / High.)

Strategic Implications

  1. The gating-vs-diffusion inversion is the defining structural feature of the 2026 frontier. The West gates the unsafeguarded tier and withholds it from all but vetted partners — the two-tier model in Fable 5 and the Two-Tier Frontier; China diffuses the safeguarded-to-state-values tier to anyone who will build on it. Both are access-governance regimes — one controls by withholding the brakes-off configuration, the other by exporting the value-aligned one. (Assessment / High.)

  2. Compute remains the one durable chokepoint — but open weights route around it for diffusion where they cannot for training. China cannot yet reliably pre-train at the frontier on domestic silicon (§2), and need not, to dominate the open-weight layer: one capable open model, given away, reaches further than any closed model can be sold. The chokepoint binds production and leaks distribution. (Assessment / High.)

  3. The cognitive-warfare exposure is real in mechanism, unproven in population-level effect. The guardrails are weight-level, default-on, reversible-but-rarely-reversed, language-differentiated, and propagating into Global South derivatives — documented mechanism; the population-scale effect is not. Track the mechanism; do not assert the outcome. (Assessment / High on mechanism; Gap on effect.)

  4. Value-alignment-by-default in the Global South’s base models is the quiet long game — more consequential than benchmark parity. Whoever supplies the base layer supplies the defaults: the refusals, framings, and analytical postures in the weights. Parity is the headline; default alignment of the world’s most-downloaded models is the durable position. (Assessment / Medium.)

Next Actions

  1. Create actor stubs for DeepSeek, Alibaba Qwen (as a distinct model-family entry), and Zhipu/Z.ai in 01 Actors & Entities/14 Corporations & Tech/.
  2. Expand the Golden Shield Project stub to connect the historical Great-Firewall censorship apparatus to its weight-level successor (censorship migrating from network to model).
  3. Author a Concepts note on “censorship-by-design / value-alignment-as-export” in 02 Concepts & Tactics/21 Information & Cognitive Warfare/, anchored to the §2.5 and §5 evidence.
  4. Cross-link this study into Sovereign AI and Civil-Military Fusion, and into the Western contrast at Tech-State Fusion in the Western Kill Chain.
  5. Open a collection requirement on any peer-reviewed end-user behavioral study of Chinese-model alignment effects — the load-bearing gap in §4; first such study materially upgrades the cognitive-warfare assessment.

Standing Gaps

  • DeepSeek V4 license text not located. MIT licensing is asserted by secondary sources; the primary license declaration was not confirmed. (Gap.)
  • No population-level cognitive-effect study. The causal chain from adoption to measurable belief/decision change is unproven (§4). (Gap.)
  • Distillation attribution unresolved. State-directed vs commercial competitive-intelligence framing of the 2026 distillation campaigns is undetermined in open source (§5). (Gap.)
  • Open-weighting intent unconfirmed. No PRC primary document frames open diffusion as deliberate state strategy; the deliberate-direction reading is Western-inferred (§3). (Gap.)
  • CloudMatrix 384 and CMP “thought-token” results uncorroborated. Huawei throughput claims lack independent MLPerf; the China Media Project directive extraction is unreplicated. (Gap.)

Key Connections

Sources

Primary / Neutral-Aggregator (High)

  • Artificial Analysis — Intelligence Index v4, artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index (composite benchmark anchor; Western frontier ~60–65, Qwen 3.7 Max 56.6, Kimi K2.6 54, V4-Pro 52, V4-Flash 47). [primary, neutral aggregator, High]
  • Artificial Analysis — DeepSeek V4 ranking, artificialanalysis.ai/articles/deepseek-is-back-among-the-leading-open-weights-models-with-v4-pro-and-v4-flash. [primary, neutral aggregator, High]
  • DeepSeek-V3 Technical Report, arxiv.org/abs/2412.19437 (MoE architecture, low activation ratio). [primary, lab, High]
  • MiniMax-M1 report, arxiv.org/abs/2506.13585. [primary, lab, High]
  • Baidu — ERNIE 4.5, ernie.baidu.com/blog/posts/ernie4.5/. [primary, lab, High]
  • Nvidia 10-Q FY2026 Q1 (SEC), sec.gov/Archives/edgar/data/0001045810/000104581025000116/nvda-20250427.htm ($4.5B H20 charge). [primary, corporate filing, High]
  • BIS — General Prohibition 10 guidance, bis.gov/media/documents/general-prohibition-10-guidance-may-13-2025.pdf. [primary, government, High]
  • BIS — AI Diffusion Rule rescission, bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens. [primary, government, High]
  • TechInsights — SMIC N+3 / no EUV, techinsights.com/blog/smic-n3-confirmed-kirin-9030-analysis-reveals-how-close-smic-5nm. [primary, teardown analysis, High]
  • Hugging Face — State of Open-Source Spring 2026, huggingface.co/blog/huggingface/state-of-os-hf-spring-2026 (41% download share, developer-share crossover, Qwen derivative counts). [primary, platform data, High]
  • OpenRouter — State of AI 2025, openrouter.ai/state-of-ai (~13% weekly token share, ~30% peak). [primary, platform data, High]
  • Stanford HAI — 2026 AI Index, hai.stanford.edu/ai-index/2026-ai-index-report. [primary, neutral aggregator, High]
  • White House — America’s AI Action Plan (Jul 2025), whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf. [primary, government, High]
  • OpenAI — gpt-oss, openai.com/index/introducing-gpt-oss/ (US open-weight counter). [primary, lab, High]

Government / Think-Tank (High)

  • NIST CAISI — DeepSeek evaluation (Sep 2025), nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks. [government, High]
  • NIST CAISI — DeepSeek V4-Pro (May 2026), nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro (~8-month lag; ~4× CCP-narrative echo). [government, Medium on lag estimate]
  • CSIS — DeepSeek Deep Dive, csis.org/analysis/deepseek-deep-dive (training-cost contestation). [think-tank, High]
  • CSIS Futures Lab — Hawkish AI / CFPD, csis.org/analysis/hawkish-ai-uncovering-deepseeks-foreign-policy-biases (foreign-policy bias benchmark). [think-tank, High]
  • USCC — Two Loops, uscc.gov/research/two-loops-how-chinas-open-ai-strategy-reinforces-its-industrial-dominance. [government commission, High]
  • CNAS (Fedasiuk) — cnas.org/publications/commentary/chinas-ai-is-spreading-fast-heres-how-to-stop-the-security-risks. [think-tank, High]
  • CSET/ETO — PLA procurement, cset.georgetown.edu/article/the-national-security-case-for-limiting-chinas-access-to-advanced-u-s-compute-evidence-from-pla-procurement-documents. [think-tank, High]
  • Brookings (Villasenor) — limits of export controls, brookings.edu/articles/deepseek-shows-the-limits-of-us-export-controls-on-ai-chips. [think-tank, High]
  • DFRLab — Digital Silk Road report, dfrlab.org/2026/02/25/china-digital-silk-road-report/. [think-tank, High]

Censorship / Empirical (High unless noted)

  • R1dacted, arxiv.org/html/2505.12625v1 (weight-level censorship survives self-hosting; ZH 99.57% vs Farsi 61.16%). [peer/preprint, High]
  • Pan & Xu, PNAS Nexus (Feb 2026), academic.oup.com/pnasnexus/article/5/2/pgag013/8487339 (language, not location, predicts sanitization). [peer-reviewed, High]
  • PromptFoo — DeepSeek censorship, promptfoo.dev/blog/deepseek-censorship/. [analyst, Medium–High]
  • Perplexity — R1 1776 (de-censoring, <1% benchmark loss), deeplearning.ai/the-batch/perplexity-launches-uncensored-version-of-deepseek-r1-ai-model. [industry, High]
  • China Media Project — Tokens of AI Bias, chinamediaproject.org/2026/02/09/tokens-of-ai-bias/ (thought-token directive extraction). [analyst, Medium — unreplicated]
  • CEPA — Chinese AI models spread propaganda globally, cepa.org/article/chinese-ai-models-spread-propaganda-globally/. [think-tank, High]

Analyst / Secondary (Medium)

  • SemiAnalysis — DeepSeek Debates, newsletter.semianalysis.com/p/deepseek-debates (efficiency is industry-wide; true-cost estimates). [analyst, Medium]
  • Interconnects (Nathan Lambert) — China open-source trajectory, interconnects.ai/p/on-chinas-open-source-ai-trajectory. [analyst, Medium]
  • MIT Technology Review — China open-source models, technologyreview.com/2026/04/21/1135658/china-open-source-models-ai-artificial-intelligence/. [secondary, Medium–High]
  • Dario Amodei — On DeepSeek and Export Controls, darioamodei.com/post/on-deepseek-and-export-controls. [analyst, interested party — containment school, Medium]
  • Anthropic — 2028 AI Leadership, anthropic.com/research/2028-ai-leadership. [corporate, interested party — containment school, Medium]
  • CFR — DeepSeek V4 signals new phase, cfr.org/articles/deepseek-v4-signals-a-new-phase-in-the-u-s-china-ai-rivalry. [think-tank, Medium–High]

ZH / State-labeled

  • AIDP full text (2017), gov.cn/zhengce/content/2017-07/20/content_5211996.htm. [primary, state, ZH]
  • CAC — Interim Measures for Generative AI Services, Art. 4, cac.gov.cn/2023-07/13/c_1690898327029107.htm. [primary, state, ZH]
  • Xinhua — Qwen ~700M downloads, english.news.cn/20260113/004b0522f987475cbf83ffc3a8d009aa/c.html. [state-aligned]
  • DigiChina (Stanford) — AIDP full translation, digichina.stanford.edu/work/full-translation-chinas-new-generation-artificial-intelligence-development-plan-2017/. [secondary, independent translation, High]

Assessment confidence: High on the capability convergence (neutral-aggregator anchored), the export-control timeline, the weight-level/reversible/language-differentiated censorship findings, the open-weight diffusion metrics, and the strategy spine. Medium on causal attribution of the MoE architecture to export controls, on Huawei’s self-reported throughput, on the open-weighting-as-deliberate-state-strategy intent reading, and on the unreplicated China Media Project directive extraction. The cognitive-warfare claim is real in mechanism, Gap on population-level effect — the load-bearing limit, held openly. Register: EN analyst (Style-Guide §2.1). PRC state sources labeled [primary, state] / [state-aligned], never [primary, authoritative]. Subject treated strictly as external intelligence.

Graph View