Hyperliquid is in discussions with Kraken's parent company about entering the U.S. market.

Insight Beating AI Flash News: Liquid AI has released the multimodal decision model d1-ultra, which scored 65 points in Artificial Analysis’s Decision Index v1.0, surpassing Jev 1.14’s 57 points to claim the top spot. Unlike models built for chatting or article generation, this decision model reads text and images to directly judge which option is more likely correct and provides corresponding probabilities. Its applications include inspecting circuit board defects, verifying that product images match their descriptions, or reviewing customer-uploaded claim photos. However, the new model’s top ranking comes with a notable tradeoff: in Artificial Analysis’s 500-question knowledge test, d1-ultra answered only ~32% of questions correctly, lower than the previous d1 generation’s 41%. Its strength lies in reducing incorrect responses. Among questions it did not answer correctly, the share of choosing “I don’t know” rose from 2% to 85%. When facing uncertain queries, it is more willing to admit ignorance rather than risk providing wrong answers — a key factor behind its significant ranking improvement. Artificial Analysis assigns 12.5% weight each to knowledge accuracy and the ability to avoid incorrect responses, with the remaining score derived from complex reasoning, tool call judgment, and safety classification. Liquid AI’s own 112-test evaluation shows d1-ultra outperforms Jev 1.14 across all eight task categories. d1-ultra is available via Liquid API and Vercel; its model weights have not been open-sourced. It charges $0.042 per million input tokens, 2.1 times the price of the previous d1 model, with no output fees. For a single image used to answer multiple questions, input costs are recalculated for each question.

According to market data from BIT (bit.com), the three major U.S. stock indexes showed mixed performance: the Nasdaq Composite fell 1.09%, the Dow Jones Industrial Average rose 0.44%, and the S&P 500 slipped 0.11%. AI-related stocks declined broadly, with individual components posting losses as follows: COHR down 5.89%, LITE down 4.06%, AAOI down 5.83%, SKHY down 5.13%, STX down 5.84%, WDC down 4.84%, SNDK down 4.29%, and MU down 4%.

Beating AI News Flash: China’s Ministry of Commerce and Ministry of Science and Technology plan to include 23 AI technologies in export controls, covering large language models (LLMs), AI agents, computing power clusters, and more. For LLMs, core technologies may be restricted if they meet any of three criteria: 1. Training compute reaches 10^26 floating-point operations; 2. Parameter count hits 2 trillion; 3. Ranks in the top 10 of global mainstream public evaluations. The scope of controls includes model weights, underlying source code, architecture design, and training optimization technologies. However, this does not mean Chinese LLMs cannot be open-sourced going forward. Open-source technologies already in the public domain are not bound by the new rules, and future open-sourcing of restricted technologies can go through a fast-track licensing process. The Ministry of Commerce stated that this process can basically achieve "immediate approval and export". Overseas users continuing to use Chinese models via APIs, web interfaces, or apps also fall outside the scope of this LLM technology export control. What is truly restricted is the transfer of core technologies abroad, such as handing over model weights to overseas enterprises or conducting localized deployment overseas. Some technologies involving China’s core data, national defense, and other sensitive fields are proposed to be banned from export. The new catalog is currently open for public consultation, with a deadline of November 15, while the pre-export registration requirement for related technologies has already taken effect on October 16.

Lighter has announced the official launch of its EVM development network (devnet). This EVM-equivalent rollup was built in collaboration between Lighter and Axiom, and is powered by OpenVM. It operates in sync with Lighter Core, with both designed to jointly settle to Ethereum via a single ZK proof. Developers can deploy using Foundry, Hardhat, Viem, and Ethers, and retrieve account, balance, position, order, and market data from Lighter Core. They can also place, modify, and cancel orders via smart contracts, as well as transfer assets between the devnet L1, Lighter Core, and Lighter EVM. Lighter stated that teams can apply for access via email at evm@lighter.xyz. Currently, the project is onboarding developers in batches to ensure rapid issue resolution, and will expand access to all developers in the coming weeks.

Insight Beating AI Flash News: U.S. monitoring software firm Datadog has launched its first dedicated operations and maintenance (O&M) model, Scout 1.0, built on Zhipu AI’s open-source GLM-5.3 Flash. Designed specifically for troubleshooting online system failures, Scout can query logs, monitoring metrics, and service call traces to gradually identify root causes of issues. Datadog built a simulation environment using real incident and monitoring data, then trained the model via reinforcement learning to eliminate false clues and support diagnostic conclusions with evidence. In 90 fault-troubleshooting tasks from public benchmark SREGym, Datadog’s self-test of Scout’s highest inference tier completed 80 diagnostic and repair tasks, tying with GPT-6 Astra xhigh and only 2 tasks fewer than GPT-6.1 Sol xhigh. The original GLM-5.3 Flash finished 66 tasks. In Datadog’s internal AlertEval evaluation, Scout’s mid-tier diagnostic quality is close to GPT-6.1 Sol high and Claude Opus 5.5 high, with an estimated inference cost 3 to 8 times lower. Compared to GPT-6 Astra medium, the cost is roughly 19 times lower. Scout has been integrated into Datadog’s Bits Investigation and is currently available in a closed preview to a small number of customers. At this stage, it mainly focuses on identifying fault causes and proposing recovery suggestions; automatic repair remains a future R&D goal.

Beating AI Insight News Brief: US AI firm Reflection has uploaded the weights of its first large language model Beam 1.0 to Hugging Face, fulfilling its October 5 announcement. The model uses a Mixture-of-Experts (MoE) architecture, has a total of 501 billion parameters, with around 23 billion activated per inference, and supports a 1 million-token context window. Three versions—BF16, FP8, and NVFP4—are released simultaneously under the Apache 2.0 license, allowing commercial use and secondary development. Reflection also open-sourced its accompanying programming agent framework Mirror CLI, which supports tool calling, parallel task execution by multiple agents, and custom model interfaces. The company additionally launched DFlash2, a model accelerator that generates multiple candidate tokens in advance for validation by the main model. The vLLM team optimized Beam’s underlying code using the programming agent: on 8 GB200 GPUs, the per-step computation time for single-request decoding dropped from 8.95 milliseconds to 3.30 milliseconds, reaching 2.71 times the speed of the initial version. The release is backed by over 40 partners including AWS, Google Cloud, and NVIDIA. A recent independent evaluation by Artificial Analysis scored it 30 points, ranking second among open-weight models in the US, trailing only Meta’s Muse Spark Open (39 points). However, it falls behind Chinese models such as DeepSeek V4 Flash (39 points) and Zhipu GLM-5.3 (45 points). Reflection is valued at approximately $25 billion. On October 10, the Financial Times reported that NVIDIA, which has already invested around $800 million, is discussing further capital injection or an acquisition of Reflection.