DeepSeek Archives - The AI Prism

China’s Open-Weights Model Strategy and the Global AI Adoption Race

The AI Prism Admin — Wed, 05 Aug 2026 14:00:00 +0000

The open-weight race is no longer a sideshow

For most of the past decade, the story of advanced AI was a story about who could train the single best closed model. That framing now obscures the more important contest: who supplies the weights the world actually runs. You can download a frontier-grade Chinese model tonight and fine-tune it on your own hardware, an option no US frontier lab offers at parity.

This shift is not a footnote. It is the structural change that explains why a Qwen or a DeepSeek now sits underneath products built by companies that will never appear on a public leaderboard. The center of gravity in AI is moving from the model that scores highest to the model that is cheapest to deploy at scale.

The moat was never in the model itself. As one observer notes, the durable advantage lives in the enterprise services wrapped around a model — the contracts, the integrations, the quality-of-life features — not in the weights (werd.io, 2025). Open release turns a US compute disadvantage into a distribution advantage and commoditizes the very layer where American cloud vendors earn their margin.

Measuring adoption is inherently hard, and download counts are an imperfect proxy for real deployment. Yet the direction of the curve is unambiguous: the open layer is being supplied, at scale, from labs that Washington does not control, and that fact is now shaping policy rather than the other way around.

What “open weights” actually buy you

An open-weight model publishes its parameters, so you can run it on your own servers, modify it, and keep your data inside your own trust boundary. That autonomy is the entire point for teams that cannot or will not route sensitive workloads through a foreign API. Closed providers sell access; open providers hand you the model.

The practical difference shows up in cost, control, and the freedom to keep iterating without a vendor’s permission. When you own the weights, a price hike or a policy change at the lab cannot switch off your product. That resilience is why adoption has compounded rather than stalled, and why regulated industries such as healthcare and finance lean toward self-hosted open models.

Open does not mean risk-free. Running a model locally, on a trusted cloud, or via a neutral inference provider such as Hugging Face removes most data-sovereignty concerns, but many adopters still default to the lab’s own app or API (Stanford HAI, 2025). The dependency question is real, yet it is a choice the buyer controls in a way a closed API never allows. For governments pursuing “sovereign AI,” an open model run on domestic hardware is the cleanest path to autonomy.

The download ledger: Qwen overtakes Llama

In September 2025, Alibaba’s Qwen family passed Meta’s Llama to become the most-downloaded LLM family on Hugging Face, a milestone documented in Stanford’s DigiChina brief (Stanford HAI, 2025). By early 2026 Qwen had crossed 1 billion cumulative downloads, far ahead of any Western open family (index.dev, 2026).

The geographic split is just as telling. Between August 2024 and August 2025, Chinese developers accounted for 17.1% of all Hugging Face downloads versus 15.8% for US developers (Stanford HAI, 2025). In September 2025, Chinese-base derivative models made up 63% of all new fine-tuned releases on the platform.

The breadth behind those numbers is striking. Reports indicate 8 of the top 10 open-source large models are now Chinese, and Qwen alone generated 153.6 million downloads in February 2026 — more than double the combined total of the next eight major players (index.dev, 2026). Qwen has also spawned over 200,000 derivative models, the first open foundation model to reach that scale, compared with roughly 72,000 for Google and 46,000 for Meta.

Cost is the quiet adoption engine

You do not adopt a model because a benchmark says it is best; you adopt it because it is cheap enough to ship. Chinese labs price inference at a fraction of US frontier rates, which matters most for coding and high-volume workloads where tokens add up fast. The decision is arithmetic, not allegiance.

According to aggregate reporting, roughly 80% of US AI startups now build on Chinese open models, and Chinese open models climbed from 1.2% to nearly 30% of global AI usage share within a single year (index.dev, 2026). For a cash-strapped startup, a price gap of roughly 3x below Gemini-class models and as much as 12x below top US flagships is not a detail; it is the difference between a viable product and a closed beta (index.dev, 2026).

Cost also explains the workload mix. As coding rose from about 11% of routed LLM usage at the start of 2025 to over 50% by mid-2026, Chinese models — strong and cheap on code — captured the surge (index.dev, 2026). Adoption follows the cheap, good-enough tier, and that tier is overwhelmingly Chinese. Premium reasoning remains a smaller niche where US labs still command revenue and enterprise trust.

A portfolio of labs, not a single champion

Treat “Chinese AI” as one actor and you miss the structure. The field is a portfolio: Alibaba’s Qwen for ecosystem breadth, DeepSeek for price-performance, Zhipu’s GLM for enterprise and government, and Moonshot’s Kimi for coding and tool use. Each lab pursues a different control point rather than a single national champion.

Architecture choices reinforce the strategy. Many Chinese labs lean on Mixture-of-Experts designs that squeeze more performance from limited compute, a direct response to US export controls on advanced chips (Stanford HAI, 2025). Efficiency under constraint is not a compromise; it is the product thesis. Even Baidu, long a voice for proprietary models, reversed course in June 2025 and released its Ernie 4.5 weights openly.

The ecosystem is deep, not narrow. More than a dozen Chinese organizations now release powerful models openly, from university labs to cloud giants such as Tencent and ByteDance (Stanford HAI, 2025). Zhipu’s GLM-4.5 uses multi-expert training for balanced, generalist capability, and by late 2025 Zhipu reported a tenfold overseas user surge to some 100,000 API users. Alibaba markets Qwen as an “AI operating system” with clients such as HP and AstraZeneca. The commercial logic is to seed adoption with free weights and capture the monetizable tail through cloud and fine-tuning.

The shock that moved markets

DeepSeek’s January 2025 release did more than impress researchers; it moved markets. Nvidia shed close to $600 billion in market value in a single session, the largest one-day loss in US history at the time, as shares fell 17% (CNBC, 2025). The sell-off hit much of the US tech sector and pulled down Dell, Oracle, and Super Micro alongside it.

The panic reflected a simple fear: if a lab can train a competitive model for under $6 million on export-compliant H800 chips, the compute moat looks far narrower than assumed (CNBC, 2025). Broadcom lost 17% and $200 billion the same day, a signal that investors questioned the entire spending thesis. The episode became a “wake-up call” that reshaped US policy thinking within months and pushed open weights onto the Washington agenda.

Why US frontier labs stayed proprietary

Most US frontier labs kept their flagship weights closed, betting that a capability lead and enterprise trust would outweigh the distribution advantage of openness. That bet is now under pressure as open rivals close the quality gap on all but the hardest agentic tasks. Proprietary release remains a strategic choice, not a technical necessity.

The pattern fits a broader retreat from open research among leading US labs, a trend we examined in why the hottest AI startups stopped publishing research. When the best work moves behind APIs, the open ecosystem loses both talent visibility and a training signal for the next generation of builders. The US response has been late but real: OpenAI released open-weight gpt-oss models under Apache 2.0 in August 2025, and the White House’s July 2025 AI Action Plan elevated open weights as a strategic asset for innovation and security.

Yet the US still treats its strongest models as closed by default, while China treats openness as the default for its strongest public releases. That asymmetry in release strategy, more than any single benchmark, is what is reshaping who builds on whom — and it creates a branding barrier of its own, as some US firms cannot use Chinese weights for compliance reasons regardless of quality.

Washington’s policy crossroads

The Trump administration’s AI Action Plan tightened export controls on foreign adversaries while naming open-weight models a strategic asset. The harder question is whether to extend those controls to foreign open models themselves, treating a downloadable file like a controlled export. That step would mark a sharp break from how the US has treated open software for decades.

The January 2025 Framework for AI Diffusion created ECCN 4E091 to control the weights of the most advanced closed models, but pointedly excluded open-weight releases. Senator Josh Hawley’s proposed “Decoupling America’s AI Capabilities from China Act” would bar importing any Chinese model, including open-source ones. Export-control scholars argue such blanket limits would be porous and would mostly punish domestic innovation without stopping proliferation (Just Security, 2025). The lighter-touch path they propose is model-by-model risk assessment instead of identity-based bans, coupled with independent oversight.

Startup founders push back

In July 2026, nearly 200 Silicon Valley companies — including Proton and Y Combinator’s network, organized through the new Little Tech Association — urged the administration not to cut off access to Chinese open-weight models (Politico, 2026). Their letter argues that American leadership requires both world-leading US open models and continued access to open models already available worldwide.

Their warning is blunt: a ban would not stop proliferation but would “instantly” kill hundreds of US startups that rely on cheap open weights instead of pricey US API credits (Politico, 2026). One founder estimated “there’ll be hundreds of companies that instantly die,” while a White House official said the goal should be “the lightest-touch way that doesn’t raise costs, limit access or inhibit American innovation.” The debate spilled onto Hacker News, where the story drew more than 1,000 upvotes and 800 comments (Hacker News, 2026).

The safety counterargument

Not everyone equates openness with progress. Anthropic’s CEO argues his company has never advocated a blanket ban, but urges focus on keeping powerful chips out of authoritarian hands, stopping industrial-scale distillation, and requiring safety testing of capable models (Anthropic, 2026). The concern is durable rather than partisan, and it is shared across the US national-security community.

Once weights ship, guardrails can be stripped and copies spread beyond any monitor, which is why open release creates a persistent risk that closed deployment does not. An evaluation by the US AI Safety Institute found DeepSeek models were on average 12x more susceptible to jailbreaking than comparable US models (Stanford HAI, 2025). The UK AI Security Institute makes the same structural point: openness precludes the safeguards closed developers can apply, and once weights are out the options are lost permanently. The open question is whether pre-release testing, rather than import bans, is the lighter-touch safeguard that still addresses the risk (Anthropic, 2026).

What the divergence means for the global order

The strategic conclusion is narrower than “China is winning AI.” The open model layer has been commoditized, and Chinese labs supply much of it — a distribution advantage that reaches the Global South precisely where US frontier APIs are costly or unavailable. For lower-income adopters, a good-enough open model is often the only advanced AI they can run at all.

The diplomatic framing matters. Beijing packages open model sharing and AI infrastructure support as tools for equitable, sovereign development, implicitly contrasting them with US export controls and closed models (Stanford HAI, 2025). At least 72 local government agencies across China had integrated localized DeepSeek models into governance systems by March 2025, a sign of how fast open models convert to institutional adoption. Gulf states and others are already weighing where to anchor their sovereign AI stacks, a calculation we detailed in the GCC’s AI policy.

Censorship and governance concerns travel with the models, and adopters should weigh them against the cost advantage. If adoption follows price and permissionless access, the center of gravity in AI may settle far from where the most capable closed models are trained. The open question is whether the US responds with its own competitive open models or with restrictions that accelerate the very dependence it seeks to prevent?

References

Stanford HAI & DigiChina Project. Beyond DeepSeek: China’s Diverse Open-Weight AI Ecosystem and Its Policy Implications. (hai.stanford.edu) — Qwen overtakes Llama in Sept 2025; Chinese developers 17.1% vs US 15.8% of HF downloads; 63% of new derivative models China-based; DeepSeek 12x jailbreak susceptibility per CAISI/AISI; 72 local agencies on DeepSeek; Zhipu tenfold overseas surge.
index.dev. The Global Rise of Chinese Open Source AI Models. (index.dev) — Qwen 1B+ downloads, 200,000+ derivatives, 80% of US startups on Chinese open models, ~30% global usage share, 3x-12x price gaps, 8 of top 10 open LLMs from China, Feb 2026 download spike.
CNBC. Nvidia sheds almost $600 billion in market cap, biggest drop ever. (cnbc.com) — Jan 27 2025 sell-off (17% drop, ~$600B, Broadcom -$200B); DeepSeek trained for under $6M on H800 chips; Nvidia later regained the top spot.
Politico. Startup founders urge Trump not to shut off Chinese open weight AI. (politico.com) — ~200 Silicon Valley companies via Little Tech Association letter, July 2026; “hundreds of companies instantly die” warning; Kratsios “lightest-touch” framing.
werd.io. American AI is locked down and proprietary. It’s losing. (werd.io) — open beats proprietary on infrastructure adoption; 80% startup adoption cited via a16z/Casado in The Economist; moat is in services, not weights.
Anthropic. Our position on open-weights models. (anthropic.com) — no blanket ban; focus on chips, distillation, safety testing; lighter-touch safeguards over import bans.
Just Security. Export Controls on Open-Source Models Will Not Win the AI Race. (justsecurity.org) — model-by-model risk assessment over identity-based bans; ECCN 4E091 context; export controls on open models called porous.
Hacker News. Discussion of the Politico story. (news.ycombinator.com) — 1,000+ points, 800+ comments, July 2026; signals close developer-community attention.
Understanding AI / Nathan Lambert (ATOM Project). The best Chinese open-weight models. (understandingai.org) — field map of Qwen, DeepSeek, GLM, Kimi; “Qwen alone is roughly matching the entire American open model ecosystem.”

The post China’s Open-Weights Model Strategy and the Global AI Adoption Race appeared first on The AI Prism.

Why AI’s Hottest Startups Stopped Publishing Research

The AI Prism Admin — Tue, 04 Aug 2026 19:56:59 +0000

In 2019, OpenAI published a paper about GPT-2, then withheld the full model for months over “concerns about malicious applications.” In 2022, it published detailed technical write-ups of DALL-E 2 and InstructGPT. Anthropic spent 2023 releasing one interpretability paper after another. If you built on this research, you knew exactly what you were working with.

Now? GPT-4’s report explicitly withheld the architecture, hardware, training compute, and dataset construction. And it’s only gotten quieter since. Technical reports became system cards. Open weights became a policy debate.

Here at The AI Prism, we’ve been tracking this shift, and the data confirms it: the great AI transparency reversal is real, measurable, and deliberate. Frontier labs didn’t drift into secrecy — they chose it. And the open source community, once written off as a hobbyist sideshow, rushed into the gap.

This is the story of how the most transparent research culture in tech history closed its doors — and why open source is now the only place you can actually see the work.

The Paper Mill Closed in 2023

Let’s pin the exact moment. OpenAI’s GPT-4 Technical Report (March 2023) reads like a scientific paper and behaves like a press release. Its own words: “Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.”

On Hacker News, the reaction was immediate — and brutal. “OpenAI should be called ClosedAI” became a running joke in March 2023. Critics noted the irony of a company named OpenAI refusing to disclose the size of its own model.

It didn’t change anything. Every major model since — GPT-4o, o1, the GPT-5 line — shipped with a “system card,” not a technical report. GPT-5.2’s launch in December 2025 was a blog post, a benchmark chart, and a safety card. No architecture. No data. No training details.

The pattern holds across the industry. Stanford’s Foundation Model Transparency Index ranked OpenAI in the top tier in 2023. By its December 2025 edition, the same index ranked OpenAI 6th out of 13 companies, down 14 points.

The Competitive Calculus Behind the Silence

Why did the labs close up? Start with the economics. As Ben Werdmuller put it in July 2026: “AI models, as a product in themselves, have very little moat beyond what amounts to brand loyalty and superficial switching costs.” When the model is the product, publishing how it works is giving away the recipe.

The shift tracks the money. OpenAI restructured around a for-profit arm and started selling API access by the token. Anthropic did the same. Once revenue depends on a proprietary model, a technical report is a liability, not a contribution.

By 2026, the fear has a name: open weights. Axios reported in July that OpenAI and Anthropic quietly aligned on the threat open-weight models pose “to their bottom line” — the headline said it plainly. Anthropic CEO Dario Amodei’s July 27 position paper pushed for cracking down on “industrial-scale distillation” and keeping powerful chips out of Chinese hands.

The irony wasn’t lost on Hacker News: the post drew 1,742 comments, many calling it “ladder pulling” — pull the ladder up now that you’ve climbed it. Distillation, after all, is how many labs build their own models. The Treasury Department is now investigating whether Chinese companies improperly distilled American models to build their own.

Safety, National Security, or Both?

To be fair: the labs have reasons beyond profit, and some are legitimate. Amodei’s post lays out two nightmare scenarios — authoritarian governments building more powerful AI, and capable models misused for cyber or biological attacks. “Open-weights models that don’t have dangerous capabilities are a public good,” he wrote.

His three proposed measures: no powerful chips to China, a crackdown on industrial-scale distillation, and mandatory safety testing for “all sufficiently capable models, open and closed.” That last one is genuinely even-handed — it would apply to frontier labs too.

But notice what’s missing: none of it requires publishing research. The policy asks are all about control — of chips, of distillation, of release decisions. Transparency, the value the field was founded on, isn’t on the list. OpenAI’s Frontier Safety Framework, published in December 2024, set thresholds for tracking dangerous capabilities — but how the company tests and enforces them stays internal. We dug into the wider alignment debate in our 2026 safety analysis, and the pattern is consistent: as safety frameworks mature, the underlying research gets quieter, not louder.

What the Data Says: Transparency Is Falling, Measurably

This isn’t a vibe. Stanford’s FMTI December 2025 edition — 100 transparency indicators across 13 companies — found the mean score dropped 17 points year over year, to 41 out of 100. The individual scores tell the story:

OpenAI: -14 points, falling from 2nd place in 2023 to 6th in 2025.
Meta: -29 points, from 1st to 5th — even the open-weights pioneer closed up.
Mistral: -37 points, the biggest drop among returning companies.
xAI and Midjourney: 14 points, tied for last.
Only 30% of contacted companies submitted transparency reports in 2025, down from 74% in 2024.

Stanford’s AI Index adds the structural stat: nearly 90% of notable AI models in 2024 came from industry, up from 60% in 2023. The people building the models are companies, and companies answer to shareholders first.

And here’s the twist that matters most: even the open-weight Chinese labs scored poorly. DeepSeek scored 32; Alibaba scored 26 — despite releasing weights anyone can download. Open weights and transparency are not the same thing, and the index proves it.

Open Source Filled the Gap — and Got Within One Release Cycle

While the labs went quiet, the open source ecosystem went loud. The template was set in January 2025, when DeepSeek released R1 — weights, a technical report, and a training methodology under a permissive MIT license. The paper was so complete it was later published in Nature. Pure reinforcement learning, no human-labeled reasoning traces — researchers could read it and rebuild it. Hacker News gave it 1,843 upvotes.

By July 2026, the gap is nearly gone. Mozilla’s State of Open Source AI report measured the best open model (Moonshot’s Kimi K3) at 57 points on the Artificial Analysis Intelligence Index vs. 61 for the best closed model (Claude Opus 5) — fourth overall, ahead of three of the biggest closed labs. Epoch AI puts the open frontier at 156 vs. the closed frontier’s 162: six points, about one release cycle, with overlapping confidence intervals.

The economics are brutal for the closed camp. Kimi K3 sits 3.6 points off the top at about a third of the price, and took first on LMArena’s Frontend Code Arena at 1,679 Elo. GLM-5.2, released under an MIT license, reports 62.1% on SWE-bench Pro vs. 58.6% for GPT-5.5. Thinking Machines shipped Inkling, a 975B open-weights model, in July 2026. Google keeps pushing Gemma. Hugging Face hosts over two million public models.

The usage numbers are the real tell: at the end of 2025, about a third of OpenRouter’s tokens went to open-weight models. Now the seven highest-volume models on the platform all ship open weights. For most production workloads, the open frontier already clears the bar.

The Kubernetes Lesson: Permissionless Beats Locked Down

Open source has been here before. Tobi Knaup, who co-founded Mesosphere and watched Kubernetes eat his company’s platform, wrote the definitive analogy: open weights are having their Kubernetes moment. “Once an open platform that people can customize becomes the industry’s center of gravity,” he wrote, “no single vendor can match the combined rate of innovation around it.”

The infrastructure already exists: vLLM, SGLang, llama.cpp, Ollama, and MLX — a full serving stack built by the community, no permission required. Around Qwen and Gemma, developers produce quantized weights, LoRA adapters, model merges, and runtime ports at a pace no single lab could match.

The warnings were early and ignored. Google’s leaked “We Have No Moat” memo (May 2023) told the company that open source communities were eroding its advantage. Mark Zuckerberg spent July 2024 arguing that open source AI is the path forward. The “Open source AI must win” campaign drew 1,600+ Hacker News points in June 2026.

Even the closed labs’ own ecosystem is defecting. On July 24, 2026, an open letter from Nvidia, Microsoft, Meta, and others warned against overregulating open-weight models. Startup founders, via the newly formed Little Tech Association, urged the administration not to cut off Chinese open-weight models. Even a16z partner Martin Casado’s claim that 80% of startups use Chinese models — disputed on HN but directionally telling — points the same way: the model layer is commoditizing, and value is moving up to the harness. We mapped that war in our breakdown of open vs. closed source AI in 2026.

The Open-Washing Problem: Weights Are Not the Whole Story

Before you declare victory for open source, sit with the uncomfortable part. Open weights are not open source. The Open Source Initiative’s definition of open source AI requires training code and enough data documentation to rebuild the system. Almost no “open” model meets it — the weights are permissive, the recipe is still secret.

That’s why DeepSeek and Alibaba score so poorly on transparency despite open weights. Releasing weights lets you run the model; it doesn’t tell you how it was trained, on what data, or with what safeguards. Knaup calls it out directly: most “open source” models are more accurately “open-weight.”

There’s also a long history of open-washing — marketing source-available or weight-only releases as “open.” OpenAI’s own GPT-OSS releases in August 2025 were open weights, not open research: no training data, no recipe.

Here’s what this means: the transparency reversal didn’t create two clean camps — “closed and secret” vs. “open and honest.” It created a spectrum, and most companies, including the open ones, sit closer to the middle than they admit. Tom Bedor, writing in defense of open models, still concedes the field’s terms: the arguments against open source AI are mostly weak, but the honesty gap is real on both sides.

What to Do About It

You don’t get to fix the labs’ incentives. You do get to stop building on trust alone. A few practical moves:

Benchmark open weights yourself. Artificial Analysis and LMArena give independent, current comparisons. Don’t rely on vendor charts — they measure what flatters them.
Read the card, then read between the lines. A system card is a marketing artifact with a safety section. Ask what it doesn’t say: data sources, eval construction, training compute.
Design for model-swappability. The moat is the harness, not the model. Abstract the API, keep prompts portable, and you can switch suppliers — or host open weights — without rebuilding.
Put transparency in your RFPs. Use the FMTI’s indicators as a checklist. Vendors who won’t disclose training data or eval methodology should discount accordingly.
Contribute to open evals. Terminal-Bench, SWE-bench, BrowseComp — the open eval stack is the community’s answer to opaque model claims. More contributors, harder to fake.
Watch the policy fight. Chip export rules, distillation crackdowns, and mandatory safety testing are all live debates in 2026. They’ll decide what you’re allowed to run — and from whom.

The Bottom Line. The great AI transparency reversal is real — measured, deliberate, and now embedded in the business model of every frontier lab. The research culture that built this field closed its doors in 2023, and it isn’t coming back on its own.

What happened instead is almost poetic. The open source community — the same one the labs once treated as a research pipeline — took the gap, and is now one release cycle from the frontier, at a third of the price, with the weights in hand. The labs traded transparency for a moat that the market is commoditizing anyway.

So here’s the question we keep coming back to: if the most valuable AI companies in the world won’t show their work, and open source is now six points behind — who is actually doing the science, and who is just selling trust?

References

The post Why AI’s Hottest Startups Stopped Publishing Research appeared first on The AI Prism.