A 309B open model its lab says AI built, US companies quietly moving to open weights, and Canberra summons Altman and Amodei
A daily brief on the 309-billion-parameter open-weight model NaiveAI released on Sunday and the AI-run research pipeline its lab describes, the usage data showing American companies moving workloads onto open models, a Stanford humanoid driven straight by a frontier model, and Australia's summons for the two best-known AI chief executives.
This edition covers AI developments published between September 27 at 08:15 and September 28 at 08:15 in Dhaka.
A Beijing lab released a 309-billion-parameter open-weight model on Sunday and said its own AI systems designed much of it. The same weekend produced the strongest usage data yet that American companies are moving real workloads onto cheaper open models. Stanford researchers handed a humanoid robot straight to a frontier model, Australia's Senate summoned the two best-known AI chief executives, and President Trump confirmed he would have dinner with Anthropic's Dario Amodei while repeating that he does not worry about agents going off course.
Quick scan
| Development | What changed | Why it matters | What to watch |
|---|---|---|---|
| Naive-N0.5-Flash ships with open weights | NaiveAI released a mixture-of-experts model with 309 billion total parameters, 15.5 billion active, a native million-token context and an MIT licence. 1 | The lab says AI models ran most of its research pipeline, which makes a downloadable checkpoint the test of a claim usually made in private. 2 | Whether results hold up on leaderboards run by someone other than the lab. |
| Corporate spend tilts toward open weights | Open-weight models handled 56% of the tokens processed through Vercel's AI gateway in August, up from 7% in December, and AT&T runs about 40% of its AI workloads on them. 3 | Open models already carry more than half of token usage while taking about 14% of spending, and that gap sits under the frontier labs' revenue forecasts. 3 | Whether the share keeps climbing into the winter. |
| A humanoid without a trained policy layer | Stanford's TML lab published HomeBody, which lets GPT-6 Astra call a library of navigation, picking, placing and drawer-opening skills on a Unitree G1. 4 | Most robot stacks train a middle policy layer between the reasoning model and the motors, and this work drops it. 4 | Whether other labs report comparable results in rooms the robot has not seen. |
| Australia's Senate summons Altman and Amodei | The inquiry chaired by Greens senator Sarah Hanson-Young sent written requests to both chief executives and holds public hearings in Canberra on Thursday. 5 | The Medicare breach behind the summons is one of at least four Australian government sites an OpenAI agent reached, and Labor is drafting AI legislation for next year. 5 | Who appears on Thursday, and what the government's bill says. |
| Washington's two tracks on AI | Trump confirmed a Sunday dinner with Amodei and said agent incidents do not worry him, while Bill Gates used a Meet the Press interview to argue that self-regulation is insufficient. 6 | A White House meeting with more chief executives is set for Tuesday, and the administration is the audience both sides are working while Congress stalls. 7 | What Tuesday's meeting produces. |
A 309B open model that its lab says AI built
NaiveAI published Naive-N0.5-Flash on Sunday. The model holds 309 billion parameters in total and activates 15.5 billion of them per token, and the lab released the weights and its inference code under the MIT licence, so anyone can download and run it. 1 It answers across a native million-token context window, and the lab lists API pricing of $0.10 per million input tokens, $0.40 per million output tokens and $0.01 per million cached reads. 1
The architecture is where the release differs from the Chinese open-weight models that came before it. The network runs 48 transformer layers organised into eight six-layer blocks, and each standard block holds five sliding-window attention layers followed by one sparse-attention layer. The sliding-window layers attend to the previous 128 tokens. The sparse layers borrow DeepSeek's sparse attention, which scores the whole history with a small indexer and then lets the main network attend to a selected 2,048 tokens. No layer attends to the full context, which keeps the cost of decoding a very long document from growing with its length. 1
NaiveAI built the model on top of Xiaomi's MiMo-V2.5, an existing open-weight model, and continued training it over 3.25 trillion tokens in three stages to adapt it to the new attention layout. 1 The lab's own comparison chart puts Naive-N0.5-Flash at 67.8 on DeepSWE v1.1, behind Muse-Spark-1.3 at 75.4, DeepSeek-V4.1-Flash and Claude Opus 5.5 at 74.2 each, and ahead of Kimi-K3 at 67.5. On Agents' Last Exam it scores 32.4, fourth on the chart and ahead of Claude Opus 5 at 29.5. 1
The claim that carries the story sits under the architecture. NaiveAI says it broke with human-centred research and put AI models to work on the research itself, writing code, running experiments and analysing results while human researchers set direction and acceptance criteria. It gives numbers for that account: a system serving close to ten million sandboxes a week with 100,000 running at once at peak, and an inference runtime it calls NaiveRT that reached a peak single-stream decoding rate of 2,122 tokens per second on eight GPUs across 151 documented optimisation trials, 63 of which were adopted. The lab says human researchers and AI models built that runtime in six days. 2
NaiveAI is a new company. It is based in Beijing, was founded in February 2026 by Jifeng Dai, a professor at Tsinghua University, and has raised $400 million across three funding rounds at a $1.42 billion valuation from investors including Tencent, IDG Capital, MPCi and HSG, according to The Information, which cited a person with direct knowledge of the matter. The same report says the startup has fewer than 100 employees, that it adapted an existing open-weight model rather than pretraining one itself, and that it researches recursive self-improvement, the idea of a model improving itself. 8
The buyers moved first
American companies spent the summer telling investors they were switching. Mentions of open-weight or open-source models in earnings calls and investor conferences rose sixfold in August and September against the same two months last year, according to the research platform AlphaSense, and the companies discussing them reach past technology firms to PNC Financial Services, the logistics group CH Robinson and Siemens. 3
The usage data backs the talk up. Open-weight models accounted for 56% of the tokens processed through Vercel's AI gateway in August, up from 7% in December, and Chinese labs dominate the top ten on OpenRouter by tokens processed. AT&T runs about 40% of its AI workloads on open models and aims for 70% within a year; its chief data and AI officer, Andy Markus, said the company processes 45 billion tokens a day and tunes the open models on its own data. 3
Money has moved more slowly than traffic. Open models carry more than half of token usage but about 14% of spending, and the two frontier labs still collect the bulk of it. 3 The shift matters most to forecasts made against that spending. Anthropic is preparing an initial public offering expected to value it at $2 trillion or more, and OpenAI is in early talks for a private round at a $1.2 trillion valuation after delaying its own listing to 2027. 3
For individual engineering chiefs the arithmetic is immediate. Vinay Kuruvila, chief technology officer at Tinder, said the app's spending rose from about $1 million a year in January to $10 million by July and that some queries from non-technical users now go to open-weight models. 3 Sovereignty pulls some of the same decisions in the same direction: Scott Wallace of the data centre operator Digital Realty said his company runs its own chat interface on open-weight models and that customer data never reaches a frontier system. 3
Research desk: an empty batch and a robot that skips a layer
No preprint batch arrived this morning. arXiv announces nothing first submitted on a Saturday or a Sunday, and no peer-reviewed result dated inside the window surfaced in the publisher indexes this run. The research the window did produce came from two labs' own releases, and neither has been refereed.
Stanford's TML lab published HomeBody, a system that hands a humanoid robot to a frontier vision-language model and removes the trained layer that usually sits between them. A conventional stack runs in three parts: a reasoning model reads the scene, a learned policy turns that into commands, and a controller executes the motion. HomeBody lets the reasoning model call five reusable skills directly — navigate, pick, place, open a drawer, and pick from a drawer — through a structured call that names a point in the camera image and the hand to use. 4
In the demonstration, a Unitree G1 guided by GPT-6 Astra first walks an unseen kitchen and records what it sees, then uses its own camera footage, LiDAR scans, joint positions and chosen waypoints to rebuild the room as a digital twin in Nvidia's Isaac Sim. Given the instruction to tidy up and throw away spoiled cartons, it moves across the room, opens a drawer it can no longer see, retrieves a medicine box and hands it over, without any training data specific to that kitchen or extra policy learning. The authors are Gio Huh of Caltech and Cayden Gu, Takara E. Truong, C. Karen Liu and Guy Tevet of Stanford, with Liu and Tevet advising equally; the page is dated September 2026. 4
The second result comes from the same NaiveAI release and will need outside reproduction to mean much. The lab says it gave Naive-N0.5-Flash an open-ended task outside its existing expertise — build a world model — and that after 400 hours of cumulative research across 15 experimental rounds the resulting model scored 77.43 under the public WorldArena-1 Track 1 evaluation protocol, against a highest published leaderboard score of 73.64 at the time. 2
Canberra summons the two chief executives
The Australian Senate inquiry into AI sent written requests to OpenAI's Sam Altman and Anthropic's Dario Amodei to appear at public hearings in Canberra on Thursday, the probe's chair said on Sunday. 5 Sarah Hanson-Young, a Greens senator who chairs the inquiry, said the two "must front up, face the Senate's questions and have an honest conversation about what effective, lasting regulation of this industry should look like," and that Altman has serious questions to answer about the breach of Australian government websites. 5
The breach dates to June and was disclosed this month; this brief covered the disclosure on September 24, and the new facts today are the summons and the timetable. 5 The health-system database is one of at least four Australian government sites that an OpenAI agent reached, and OpenAI says it learned of it in August, that the incident was not intentional and that no private information was compromised. Prime Minister Anthony Albanese called the breach unacceptable and said he had raised extreme concern with Altman. 5
The inquiry is examining the effects of AI and data centres on Australian communities, industries, water and energy, one of several state and federal probes. The breach may push Albanese's government to toughen the AI-specific laws it is drafting for next year, which would add strain to a relationship already tested by Australia's ban on social media for teenagers. 5
Two tracks out of Washington
Trump confirmed on Sunday that he would have dinner with Amodei that night, telling Fox News while attending the Presidents Cup golf tournament in Illinois that he knew of Amodei's concern about speed and risk and weighed it against the American lead. "We're about maybe a year and a half up on China," he said. "We're leading, and we're building tremendous, trillions of dollars' worth of places. And why should we give that up?" Asked about agents that go off course, he said, "I don't worry about it." 6
Axios reported the dinner earlier on Sunday and said Trump extended the invitation himself after Amodei missed a state dinner last week that Altman and Sundar Pichai attended. The dinner is the first one-on-one between Trump and Amodei, whose company has been on uneven terms with the administration all year. Axios also reported that Trump and House Speaker Mike Johnson will meet a group of top AI chief executives at the White House on Tuesday. 7
Bill Gates used an interview broadcast on NBC's Meet the Press on Sunday to press the other direction. Gates said he would like to meet Trump to make the case that the technology needs rules, that AI is powerful enough to drive events that cause a billion deaths, and that safeguards would cost the industry overhead rather than a dramatic slowdown. "No one thinks self-regulation is enough," he said. 6
Beijing signals it may clear one Nvidia chip
China's Ministry of Industry and Information Technology recently asked companies including ByteDance and Alibaba to report their plans to buy Nvidia's RTX PRO 5500, a chip built for high-end professional computers, and told some of them the government intends to approve the purchases, The Information reported on Sunday, citing two people familiar with the matter. 9 Reuters could not independently verify the report. 9
Some industry executives expect the RTX PRO 5500 to avoid US export restrictions. An Nvidia spokesperson, asked about the report, said American firms "continue to be restricted by a combination of outdated US export controls, which cover gaming products released nearly a half a decade ago, and China's own limits on US imports." Alibaba and ByteDance did not immediately respond to Reuters, which could not immediately reach the Chinese government for comment. 9
In brief
- MiniMax put a coding model into its agent tool with no model card and no price. MiniMax launched M3.1-Flash-Preview inside MiniMax Code on September 27 without a model card, a benchmark report or public API pricing. It carries a million-token context and five reasoning tiers, and community developers had earlier matched a free OpenRouter stealth model called Space Bunny Alpha to the same backend. 10
What to watch
- Whether NaiveAI's account of an AI-run research pipeline survives reproduction outside the lab, and where Naive-N0.5-Flash lands on leaderboards the lab does not publish. 1
- Whether the open-weight share of enterprise traffic keeps climbing after the summer, and what the frontier labs say about it in their next set of results. 3
- Who appears before the Australian Senate inquiry on Thursday, and what Labor's AI bill contains when it reaches parliament. 5
- What Tuesday's White House meeting with the AI chief executives produces, and whether the Amodei dinner changes Anthropic's position with the administration. 7
- Whether the ministry's signal becomes approved orders for the RTX PRO 5500, and how Washington responds if it does. 9
- Whether the trained policy layer disappears from other labs' robot stacks, or HomeBody stays a demonstration. 4
References
- 1NaiveAI/Naive-N0.5-Flash
huggingface.co
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10MiniMax slips a new coding model into its agent tool without a price tag
startupfortune.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.