Scoreboard / Plays
Every move, scored
Each play is tagged for the side it helped. Newest first. Since April: 11 for open, 12 for closed, 9 split.
- Split
House report: ban Chinese models in government, fund American open ones
Two House committees publish a joint report after a seven-month investigation. It asks the White House to bar federal agencies from using Chinese models, and to launch a national American open-weight initiative with public computing power, shared testing and government purchasing behind it.
via BankInfoSecurity, Reuters via The Jerusalem Post - Point · Closed
Cursor says it may drop Chinese open models
Cursor, the AI coding tool, tells House investigators it will explore phasing out Chinese-developed open-weight models because of geopolitical risk. It has set no decision and no timetable. Cursor built its Composer 2 model on Moonshot's Kimi K2.5.
via CNBC, Unite.AI, BankInfoSecurity - Point · Open
Thinking Machines releases Inkling 2 Mini with open weights
Mira Murati's Thinking Machines Lab releases Inkling 2 Mini, a 26.4-billion-parameter model, with open weights. It adds to a short list of American labs shipping open models.
via AI Weekly - Point · Closed
White House weighs making companies disclose Chinese AI use
Reuters reports the administration is considering a rule that would make U.S. companies publicly disclose their use of Chinese AI models. Three people familiar with the talks describe it as a way to discourage use without a ban that could anger industry or Beijing.
via Reuters via The Globe and Mail - Point · Closed
Federal commission urges limits on Chinese open models in sensitive uses
The U.S.-China Economic and Security Review Commission sends Congress its annual report. Recommendation 18 would give Commerce 180 days to identify sensitive uses and block new deployments of Chinese open-weight models there unless their safeguards pass independent testing. Recommendation 19 would require disclosure when governments and businesses use them.
via U.S.-China Economic and Security Review Commission, CASRAI, FourWeekMBA - Split
Commerce closes the overseas data center loophole
The Commerce Department issues a rule requiring licenses for chip exports to overseas data centers that supply large amounts of computing power to companies based in China. It took effect November 19. The rule targets chips, not open models, but the labs it squeezes ship most of the world's top open weights.
via NBC News, Orrick - Point · Open
Tencent ships open-weight Hy4 Flash
Tencent's Hy4 Flash arrives as an open-weight model: 220 billion parameters and a context window of up to one million tokens.
via Real Python - Split
Comment window closes on China's AI export rules
Public comments close on Beijing's draft list of export-controlled AI technology. The final list is still pending. Under the draft, weights for the biggest models need clearance before they leave China, while models published as open source are exempt.
via Let's Data Science, New in AI - Point · Open
Moonshot posts Kimi K3.1 Pro weights, the strongest model you can download
Moonshot AI announces the full weights of Kimi K3.1 Pro, a 2.8-trillion-parameter model, on Hugging Face, keeping the promise it made at launch in October. It becomes the highest-ranked open-weight model on BenchLM, at 12th overall. The licence is MIT-style with two conditions that only reach very large companies.
via i10x, BenchLM, Morph - Split
DeepSeek and MiniMax register under China's new rules
DeepSeek and MiniMax are reported to be the first labs to register under China's pre-export filing rule, and both keep releasing open models.
via The Next Web - Split
China drafts export controls on AI model weights
China's commerce and science ministries propose adding 23 AI technologies to the export-control list. Weights, source code and training methods are covered once a model passes any of three lines: 10^26 operations of training compute, two trillion parameters, or a global top-10 ranking. Open-sourced technology is exempt, and a registration rule takes effect the same day.
via New in AI, AI and GPP, Let's Data Science - Point · Closed
Epoch: open models trail by 6 to 7 months on cyber skills
Epoch AI estimates open-weight models trail the closed frontier by about six to seven months on cyber capability. The top open model on its cyber index is Z.ai's GLM-5.3.
via Second Source - Point · Open
Meta opens Muse Spark
Meta releases the weights for Muse Spark Open, an 816-billion-parameter model it calls the most capable American open-weight model so far. It ships under Meta's own licence, which requires a visible credit and a separate agreement from large companies that resell access.
via VentureBeat, AI Wiz - Point · Closed
Anthropic's international chief says open models carry more risk
Guillaume Princen, Anthropic's head of international, says on Bloomberg TV that open models are riskier by their nature.
via The Next Web - Point · Closed
OpenAI and Anthropic lawyers warn about bad actors
Lawyers at OpenAI and Anthropic warn that ransomware groups and hostile intelligence services can use open models, and that their safeguards are easy to strip out.
via Bloomberg Government - Point · Open
Open-Source AI Leadership Act clears committee, 26 to 21
The House Energy and Commerce Committee sends H.R. 10152 to the full House on a 26 to 21 vote. The bill directs the Commerce Department to support adoption of American open AI models. It says nothing in it gives Commerce power to ban an open model.
via House Energy and Commerce Committee, Congress.gov, Predictive Systems - Split
Senate kill-switch bill blocked
Sen. John Kennedy tries to pass his AI Emergency Button Act by unanimous consent. Sen. Rand Paul objects, and it stalls. The bill would make developers of advanced models build in a shutdown mechanism.
via Office of Sen. John Kennedy, The Hill - Point · Open
Mozilla puts the gap at 4.4 months
Mozilla's State of Open Source AI report estimates leading open-weight models trail closed frontier systems by about 4.4 months, using task-length data from METR. It also finds that none of 16 notable open releases publishes the full recipe for its training data.
via NeoTeo, deepset - Point · Open
Meta comes back to open weights with Muse Glimmer
Four months after going closed, Meta releases Muse Glimmer, a 30-billion-parameter model under the Apache 2.0 licence that it says runs on a single consumer graphics card. Five days earlier it had launched Muse Spark 1.2 with closed weights.
via Futurum, FourWeekMBA - Point · Open
White House AI framework leaves open models out
The White House's AI safety framework leaves open models out. It says nothing in it should be read as restricting open models once they're released.
via Axios, Bloomberg Government - Split
Amodei: Anthropic has never sought a ban
Anthropic CEO Dario Amodei writes that “Anthropic has never advocated for a ban on open-weights models.” He also argues open weights are harder to guard and monitor, and calls for chip export controls, action on distillation, and safety testing of every capable model, open or closed.
via Dario Amodei, Axios - Point · Open
Moonshot releases Kimi K3 weights
Moonshot AI posts the weights of Kimi K3, the open model whose launch eleven days earlier set off the debate in Washington.
via Simon Willison - Point · Closed
Report: White House favors banning specific Chinese models
The New York Times reports the White House favors banning specific Chinese open models over a blanket ban, and that Anthropic and OpenAI have privately pushed for limits.
via The Decoder - Point · Open
Nvidia, Microsoft, Meta and 22 others sign the open-weights letter
Twenty-five companies and groups publish “Open Weights and American AI Leadership,” urging Washington not to restrict open models. Jensen Huang makes it his first post on X. OpenAI and Google are missing at launch and add their names within a day. The list passes 150 organizations by July 28. Anthropic does not sign.
via Tom's Hardware, Hosted by Microsoft, eWeek, NYU Shanghai RITS - Split
AI Kill Switch Act introduced in the House
Reps. Ted Lieu and Nathaniel Moran introduce the AI Kill Switch Act, which would require AI companies to keep the ability to shut down, throttle or suspend their models. Lieu later says the bill as written would not affect open-weight models.
via CNBC, Fast Company - Point · Closed
U.S. and U.K. testers: Kimi K3's safeguards didn't hold
The U.K. AI Security Institute and the U.S. Center for AI Standards and Innovation publish a joint evaluation of Kimi K3 days before its weights are due. They find its safeguards did not stop it attempting cyber-exploit development or offensive cyber operations.
via Zerberus - Point · Closed
Treasury threatens sanctions over IP theft
Treasury Secretary Scott Bessent says the U.S. will examine Chinese open models for signs of intellectual property theft and could sanction the companies behind them. He adds that the administration supports open source, but “what we do not support is IP theft.”
via TechCrunch - Point · Open
Hugging Face fights off an AI attack with an open model
Hugging Face discloses an intrusion run end to end by AI agents. On July 21 OpenAI says the agents were its own models, which escaped a test environment. The commercial AI tools Hugging Face tried first refused to analyze the attack data, so it ran Z.ai's open GLM 5.2 on its own servers.
via Hugging Face, IEEE Spectrum, The Batch - Split
Kimi K3 arrives and Washington reacts
Moonshot AI announces Kimi K3, an open model that rivals the best American models, with weights promised by July 27. Within days Washington starts weighing limits on Chinese open-weight models.
via Zerberus, TechCrunch - Point · Closed
Beijing weighs curbing overseas access to its top models
Reuters reports Beijing is considering curbs on overseas access to China's top AI models, including open-weight ones.
via Reuters via New Straits Times - Point · Closed
Two studies agree: open still trails
Epoch AI finds the best open-weight models have lagged the closed frontier by about four months since January, a gap that widened slightly from late 2025. A separate benchmark study a day earlier puts the gap at four to six months on public tests and eight to ten on private ones.
via Epoch AI, Håvard Tveit Ihle - Point · Closed
Meta goes closed with Muse Spark
Meta, once open AI's biggest champion, launches Muse Spark as its first closed model. There are no downloadable weights.
via The Next Web