model

According to Bloomberg, the AI model evaluation platform Arena completed a new round of financing of $200 million with a valuation of $3.1 billion, led by Lightspeed Venture Partners and Khosla Ventures, with participation from Salesforce Ventures and Dell Technologies Capital.Arena originated from the Chatbot Arena research project at the University of California, Berkeley, and is known for its public large model rankings, achieving an annual revenue of $100 million in June of this year.

Mike Selig, the Chairman of the U.S. Commodity Futures Trading Commission (CFTC), stated in an interview with CNBC that there will be no "race to the bottom" on the leverage issue. The CFTC will apply the mature margin models and risk control measures that have long been used in the derivatives market to maintain market integrity in the cryptocurrency market.

Microsoft stated that the new NVIDIA Nemotron model will be able to run on a single laptop and PC.

Microsoft (MSFT.O) has launched an artificial intelligence coding model that can run on personal laptops.

The co-founder and CEO of the decentralized perpetual contract trading platform Hyperliquid, Jeff Yan, stated during a fireside chat at the Token2049 conference in Singapore that traditional wealth creation opportunities on Wall Street, such as company stocks, are essentially inaccessible to the investing public before they are listed on exchanges, causing retail investors to miss out on the most significant price increases prior to listing. He noted that certain assets are tradable by only a few people during most of their growth phases, and by the time the public can trade them, the growth has already been captured by a privileged few, making this wealth creation model unsustainable.Yan stated that Hyperliquid's main mission is to expand the accessibility of wealth creation opportunities and encourage more people to participate in the financial system, with income being merely a byproduct. He mentioned that Hyperliquid's success partly stems from its perpetual contracts having no expiration date, which reduces the number of decisions traders need to make and prevents liquidity fragmentation. According to data from DefiLlama, Hyperliquid generated $72 million in revenue over the past 30 days, ranking third among protocols in terms of revenue.Blockchain asset management company Pantera stated in July that perpetual contracts, due to their structural advantages, could become one of the dominant trading tools in the global financial sector, and Hyperliquid has demonstrated the potential of blockchain infrastructure to challenge traditional markets. Jeffrey Sprecher, CEO of the Intercontinental Exchange (ICE), the parent company of the New York Stock Exchange, called on regulators to create a fair competitive environment for the launch of 24/7 on-chain perpetual contracts. In March of this year, the New York Stock Exchange partnered with the tokenization platform Securitize to advance blockchain-based stock trading infrastructure.

Google released its latest image generation and editing model, Nano Banana 2.1, on October 6, and it has been launched in the Gemini app, AI mode of Google Search, Google Ads, as well as developer tools like Google AI Studio, Flow, and Stitch. This model can generate images from text descriptions and edit existing photos.Google stated that this upgrade focuses on three aspects: visual design, mask-based editing (where only the marked area changes), and subject consistency, meaning that characters or objects retain their original features after multiple edits. In the overall preference test for text-to-image generation, Nano Banana 2.1 scored 1050 ELO points, higher than Nano Banana 2's 990 points and Nano Banana Pro's 935 points. The model can handle up to 14 reference images simultaneously, track 4 characters and 10 objects, and output up to 4K resolution with an aspect ratio of up to 8:1. Developers can also set the thinking time before generation and enable fact-checking based on Google Search and Google Image Search.When called through the Google Developer API, the standard 1K image is priced at $0.0336, about half of Nano Banana 2's $0.067; the 4K image is $0.0756, while the previous generation was $0.151, and bulk tasks can enjoy a 50% discount. The above performance data comes from Google's own testing.

Anthropic announced the expansion of access to its state-of-the-art AI models, allowing verified institutions to use models such as Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1 for cybersecurity testing.U.S. government agencies, financial institutions, and large software vendors participating in the "Project Glasswing" project will gain open eligibility, and newly joined institutions must be jointly reviewed by Anthropic and the U.S. government.

French artificial intelligence startup Mistral claims that its model ranks among the best open-weight models in the world, trained using approximately 4,000 NVIDIA Grace Blackwell GPUs.

On the 5th local time, the American artificial intelligence startup Reflection AI, supported by NVIDIA, released its cutting-edge open-weight model "Beam." The company stated that the performance of this new model is comparable to mainstream open-weight models.

Cryptocurrency analyst Darkfost stated that over the past decade, the volatility of Bitcoin prices around the "Power Law" trend line is gradually decreasing, and market cycle fluctuations are converging. The oscillation indicator peak has dropped from +169 in 2018 to +102 in 2025, with its daily standard deviation decreasing from 4.81 to 2.47, a decline of about 49%.Darkfost added that as of the current model's oscillation indicator at +34.5, the Bitcoin price is approximately $84,700. According to the model parameters for that day, an oscillation indicator of +100 corresponds to a Bitcoin price of about $157,800, which is approximately 86% higher than the current price. He also pointed out that $157,800 is not a fixed target price but rather a reference level calculated based on the current regression parameters. If the Power Law trend line continues to rise in the future, this reference price will also increase accordingly.

Vitalik.eth posted on Farcaster revealing that he is conducting a personal experiment to generate personalized diet and exercise recommendations using personal health and travel data by combining local models with remote cutting-edge models while protecting privacy.The system employs a three-layer privacy protection architecture:Identity Layer: A local model (Qwen 3.8B) replaces the user to construct query requests, avoiding the exposure of writing style that could reveal identity.Payment Layer: Uses zkAPI to hide payment information.Network Layer: Hides IP addresses through Tor.The system is currently operational, but Vitalik pointed out three major shortcomings: Tor is inefficient and has high latency in unlinking requests; the local model runs at only 20-30 TPS, and a smooth experience requires over 100 TPS; the stricter the data protection, the more limited the assistance that the remote model can provide. The relevant code has been submitted to the Ethereum zkAPI repository.

YouTuber PewDiePie stated in his latest video that his account was banned twice by OpenAI while training the small local AI model Ajax. He mentioned that he originally hoped to distill reasoning tokens from OpenAI's flagship model GPT-5.6 Sol, released on July 9, and find a study that utilized OpenAI's own API to crack encrypted reasoning content. He claimed that he did not actually crack OpenAI but was still banned, and after appealing, his account was restored after some time.After the account was restored, he used Sol again to generate seed data and was subsequently banned for the second time. OpenAI's terms of service prohibit the development of competing models using its outputs. Ajax is a 9 billion parameter model built on Alibaba's open-source model Qwen 3.5, which will run in PewDiePie's free self-hosted AI application Odysseus launched in June, and it is not yet available for download.To prevent Ajax from being censored, he used the open-source tool Heretic to remove the model's built-in refusal mechanism, a technique known as abliteration. He stated that Ajax suffered a bit of brain damage as a result, but it has about a 90% success rate in tasks like organizing the inbox and browsing the web, and he continues to train it using the GRPO method. He mentioned that when choosing which refusal behaviors to remove, he draws the line at harming others or harming himself.

AI startup Tavus released a new model called Griffin on October 1, referring to it as the first "human interaction model" that can understand and generate face-to-face conversations, paying attention not only to wording but also to expressions and pauses. Tavus stated that in tests, after a one-minute video call with Griffin-Lite, 26 out of 54 participants (48%) believed the other party was a real person; its previous generation system only convinced 1 out of 41 participants.Participants were told they would have a one-minute video call with another person to discuss things they were looking forward to this year, and only at the end of the call were they asked if they suspected the other party was not a real person. Tavus noted that those who became suspicious typically noticed within 20 seconds. This result comes from Tavus's own research page, where participants were recruited through a channel they call an independent research platform. Community comments on X pointed out that these results have not been independently verified and do not meet standard protocols.In the NVIDIA VideoFDB benchmark, Tavus claimed Griffin-Lite ranked first, with a generation dimension score of 3.83, the next best system scoring 2.8, and human reference scoring 3.92; the perception dimension score was 3.73, with the strongest baseline at 3.44 and human reference at 4.2. Tavus stated that NVIDIA independently conducted this evaluation. Griffin supports full duplex, allowing it to listen, see, and speak simultaneously, with an average audio-video delay of 0.43 seconds on the NVIDIA H100 chip. Tavus completed a $40 million Series B funding round led by CRV in November 2025.

The liquidity pledge protocol Lido announced an update to the EarnUSD fee model, which will reduce fixed fees and make fees more directly linked to the performance of the treasury. Previously, the EarnUSD fees were more platform-oriented, consisting of a 1% management fee plus a 10% performance fee.According to the announcement, the new model caps the management fee at 0.5% plus a 20% performance fee, with the initial rate set at a 0.2% management fee plus a 15% performance fee. Lido stated that after the management fee is reduced from 1% to 0.2%, the costs of holding EarnUSD will be lower when returns are low, and the higher performance fee will make the charges more aligned with actual returns. Relevant fees will be displayed on the EarnUSD treasury interface.

According to The Block, the European Central Bank has proposed three models for introducing central bank digital currency on the blockchain, including direct issuance, establishing bridging mechanisms, and implementation through private intermediaries.These models can provide settlement support for transactions such as tokenized securities, bank deposits, and stablecoins on distributed ledger technology networks.

According to Rest of World, Alibaba launched the open-source model platform ModelScope in 2022, and OSChina launched MoArk in 2023, providing model hosting, testing, and customization services for Chinese developers behind firewalls. In 2023, Chinese regulators blocked Hugging Face without disclosing specific reasons; reports indicate that the government simultaneously allowed artificial intelligence laboratories to bypass the blockade to share Chinese models globally. ModelScope currently hosts over 170,000 models and stated in March that it has 170,000 models and 250 million users; MoArk hosts about 20,000 commonly used models, both still below Hugging Face, which hosts over 3 million open models.OSChina CEO Xu Yong stated that not everyone can always use a VPN, and China needs to establish an independent artificial intelligence ecosystem that serves Chinese users, claiming that China is developing an independent ecosystem faster than during the internet era. He also mentioned that MoArk has deployed engineering teams to ensure that models can run on different types of Chinese chips, making it their mission to ensure mainstream open models are compatible with Chinese chips. Both platforms offer some services for free, with additional computing power and premium features charged; ModelScope also hosts university hackathons and has opened co-working and event spaces for AI entrepreneurs in Hangzhou.Headquartered in New York, Hugging Face is a major platform for sharing open models and datasets. In September, Nvidia announced its acquisition of the company for $12.9 billion and noted in regulatory filings that regulators might prohibit sharing Chinese models on the site.

OpenAI has launched the GPT-6.1 Sol model, whose performance is close to Astra, but the standard token price is only one-fifth of Astra.