Dutch central bank approves Qivalis' electronic money institution license, bringing euro stablecoin issuance a step closer.

Bitcoin treasury firm Strive filed an 8-K document today disclosing that it added 2,822 Bitcoin to its holdings between October 12 and 13, at an average price of approximately $83,266 per coin. The acquisition cost $235 million, bringing its total Bitcoin position to 35,600 coins.

Tether AI Research has released an open-source multilingual automatic speech recognition (ASR) model named QVAC TranscribePsy, which has around 254 million parameters and can run fully on-device—no audio upload to the cloud required. The company noted the model can process audio from smart glasses connected via Bluetooth or Wi-Fi in near real time, and also operate on onboard processors of humanoid robots. On the Core ML version for iPhone 17 Pro Max, transcribing 10 minutes of audio takes roughly 1.65 seconds. According to its published benchmark data, TranscribePsy achieves a word error rate (WER) of 8.02% on the Open ASR Leaderboard, 6.58% on Multilingual LibriSpeech, and 21.56% on Common Voice. Among six tested smartphones, it is the fastest in speed, approximately 2.6 times faster than the second-fastest model. The model supports 25 European languages for laptops and 8 European languages for mobile phones, with model weights and builds in formats including GGUF and Core ML released under the Apache 2.0 license. Tether CEO Paolo Ardoino stated that he hopes advanced speech AI can run locally on users’ own devices, ensuring data does not leave the device.

Beating AI News Flash: U.S.-based robotics AI firm Physical Intelligence (Pi) has launched the beta version of its Pi API, opening access to its foundational robotics model to more enterprises. Developers only need to upload robot operation data to fine-tune the model on Pi’s cloud computing power, then control their own robots via the API. There’s no need to train models from scratch or build complex training and inference infrastructure independently. Pi showcased real-world use cases from multiple partners, including AgRobotics’ strawberry-picking robots, autonomous loaders for self-driving trucks, and robots for food processing and server assembly. Among these, startup Almond used the Pi API to get its dual-arm robot to learn polishing the parts required for its own assembly after just 4 days of iterative training and testing. Pi had previously open-sourced robotics models such as π0 and π0.5, but developers still had to deploy and train them on their own. This time, the company has further packaged model fine-tuning, computing power, and inference services into an API, aiming to become a shared intelligent foundation for various robots. However, test access is still by application only, and pricing has not been announced yet.

DeFi lending protocol Spark has disclosed that its net protocol returns exceeded $4 million in the third quarter of 2026. In the first half of the year, DeFi lending spreads narrowed significantly, with lending capital profitability turning negative in the second quarter. After driving growth for Spark Savings USDT in Q3, related returns rebounded above capital costs. During the same period, Spark partnered with Uniswap and Sky to launch a stablecoin swap layer, which went live with $150 million in USDS liquidity, offering stablecoin swap services to banks, fintech firms and payment companies, and recorded over $14 billion in trading volume in its first month.

Morgan Stanley released its third-quarter financial results, reporting net revenue of $20.34 billion, exceeding the consensus estimate of $19.89 billion. The firm posted earnings per share of $3.03 for the quarter, while net interest income came in at $3.25 billion, beating the forecast of $2.78 billion.

Dongcha Beating AI News Flash: Finding files on Mac typically requires remembering names or keywords. If you only recall a zebra appearing in a video segment, or a podcast episode discussing sleep, traditional file search tools are of little help. Independent developer Abdur Rahim’s open-source tool DigUp solves this by enabling direct content-based file searches. Type in "zebra in video" and it will jump to the exact frame where the zebra appears; search for "segment discussing sleep" and it will pinpoint the relevant timestamp in the podcast. It can also search photos, PDFs, and even open the exact page of a specific clause in a contract. DigUp leverages Google DeepMind’s EmbeddingGemma 2 model, released on October 6. This model converts text, images, audio, and video into comparable numerical features, allowing users to find corresponding visuals or audio based on a text description. DigUp pre-indexes selected folders, and during searches, it can also precisely match details like invoice numbers and error codes. All indexing and search processes run locally on Mac, eliminating the need to upload private files to the cloud. Free and open-source, DigUp has earned over 3,000 GitHub stars in less than a week since its launch. It currently only supports Macs with Apple’s custom silicon, requiring macOS 14 or later. First-time use requires downloading an approximately 865 MB model and spending time building the index.