「$10M Profit Whale」 Suspected of Accumulating 6000 ETH in a Transaction, with an Average Price of $1582

Insight Beating AI News Flash: Researchers from Stanford University and NVIDIA Research have open-sourced CLM-8B, a System One model in the same category as Jev, designed specifically to help AI agents make quick judgments and selections. CLM and Jev serve the same core function: both enable agents to skip long text generation and directly choose from a set of candidate actions. CLM’s key distinction lies in its underlying architecture: it separates the calculation of "current state" and "candidate actions", allowing fixed actions to be precomputed and reused, resulting in faster performance in scenarios requiring sequential decision-making. In official tests, CLM-8B performs on par with Jev overall, with latency reduced by up to approximately 9 times. The team also tested CLM’s ability to select answers for coding agents: Opus 5 and Fable 5 first generate multiple candidate solutions, then the fine-tuned CLM picks the optimal one. On DeepSWE, CLM boosted Opus 5’s single-pass success rate from 73.7% to 81.6%; on Terminal-Bench 2.1, it increased from 84.0% to 87.6%. CLM does not generate code itself; the test evaluates its ability to select superior solutions from existing candidates. CLM-8B is built on a frozen Qwen3-8B, with only state and action projection heads trained additionally. Its training data includes approximately 60 million question-answer pairs, 30 million hard negative samples, and around 1 million agent trajectories. The code and model weights are both open-source, released under the Apache 2.0 license.

New York Fed President Williams said today that U.S. inflation remains significantly above the Federal Reserve’s 2% target, and the central bank still has “a lot of work to do” to curb inflation amid high energy prices and strong demand driven by AI investment. He called another interest rate hike this year a “reasonable” expectation, but stressed that future policy decisions will depend on economic data. Williams noted that the U.S. economy has shown resilience after major shocks, but persistent energy price pressures and strong demand from AI investment have raised inflation risks. He pointed out that U.S. inflation has exceeded the Fed’s target for five consecutive years. The Federal Reserve unanimously raised interest rates by 25 basis points last week, lifting the federal funds rate target range to 3.75%-4%. The latest dot plot shows 16 of 18 officials expect at least one more rate hike before the end of 2026. Williams said each future policy meeting will be decided based on the latest economic and inflation data. Markets have also increased bets on an October rate hike: CME FedWatch data shows that as of September 24, the market’s probability of a Fed rate hike in October is around 70%, up from about 54% the previous day. In addition, Williams noted that AI investment has become a new variable on the demand side. While AI investment could drive productivity growth in the coming years, its actual contribution to overall productivity remains limited for now.

The European Banking Authority (EBA), in its response to the European Commission’s targeted consultation on the Markets in Crypto-Assets (MiCA) framework, is calling for crypto lending to be incorporated into the EU’s MiCA regulatory regime. This includes cases where crypto asset service providers facilitate client access to decentralized finance (DeFi) lending protocols. The EBA recommends the European Commission conduct a cost-benefit analysis of legislative changes, add intermediated crypto lending to MiCA’s list of regulated services, and potentially impose specific compliance requirements and oversight measures. It also proposes relevant rules for crypto firms that offer clients access to DeFi lending protocols. Potential measures outlined by the EBA include suitability tests for users, leverage limits, enhanced disclosure requirements, possible restrictions on lending access involving MiCA-authorized asset reference tokens or e-money tokens, and a certification system for DeFi lending protocols. The EBA notes that crypto lending continues to grow across the EU; prior research shows lending activities exist in at least 16 member states. Easier access to DeFi via crypto firms and AI tools is increasingly blurring the line between centralized and decentralized finance.

Binance announced in an official statement that it will gradually add Hyperliquid (HYPE) to its services including Simple Earn, Buy Crypto, Convert, VIP Loan, and Margin starting September 24. HYPE’s flexible product will launch on Simple Earn at 19:00 UTC+8 on September 24, with subscriptions opening immediately. Within one hour of HYPE’s spot trading going live, users can purchase HYPE via payment methods such as bank cards, Google Pay, and Apple Pay, and trade HYPE with assets like BTC and USDT on Binance Convert with zero fees. Additionally, HYPE will become an eligible borrow asset for VIP Loan within one hour of its spot listing. Binance’s margin trading will add HYPE as a borrow asset at 19:00 UTC+8 on September 24, and launch HYPE/USDT and HYPE/USDC cross-margin and isolated margin trading pairs; portfolio margin will also support HYPE and the above-mentioned trading pairs simultaneously. Binance reminded that newly listed tokens usually have high volatility, so users should pay attention to risk management.

Beating AI Express News: Apple has open-sourced LensVLM-9B, a vision-language model purpose-built for processing long documents, trained on Qwen3.5-9B-Base. When handling long documents, the model first compresses the entire document into low-resolution pages for a quick scan to locate relevant pages, then reads the original text or high-definition images on those pages—eliminating the need to load the full document into the model’s context window. Per the paper’s tests: For a 100-page document, reading the full text directly requires 51,273 tokens, while LensVLM uses just 8,090 tokens. The corresponding KV cache drops from approximately 1.6GB to 253MB, an 84.2% reduction, with no significant accuracy loss after compression. Among 7 document question-and-answer tasks, direct full-text reading yields an average accuracy of 72.4%, while LensVLM hits 68.9% at roughly 4.3x compression. Compared to the approach of converting text to images for model recognition (at around 5x compression), LensVLM lifts accuracy from 31.3% to 68.9%. That said, the model is currently slower: After locating the relevant page, LensVLM must call an additional tool to read the original text, requiring two consecutive inference runs. The paper notes each answer takes approximately 17 seconds; without the compression scheme, directly inputting the full document as text into Qwen3.5-9B takes about 8 seconds to generate a response.

Bitget Onchain, its on-chain trading service, has now integrated Robinhood Chain, enabling users to directly trade all assets in the Robinhood ecosystem using the stablecoin U (United Stables) from their spot wallets. This integration further expands Bitget UEX’s panoramic ecosystem layout, delivering a more efficient on-chain trading experience for users. As of now, Bitget Onchain has integrated six major networks: Robinhood, MORPH, ETH, SOL, BSC, and Base. Users can directly access millions of on-chain assets and US stock trades via their spot wallets, and leverage auxiliary tools including AI smart money copy trading and on-chain limit orders to achieve intelligent analysis and trade execution.