分析师：山寨币狂热出现首个警示信号，上涨动能可能正减弱

According to YuEmber monitoring, a whale or institution that has held ETH for three years realized further profit-taking of 30,825 ETH (worth approximately $83.03 million) 9 hours ago. Over the past week, the entity has transferred a total of 112,053 ETH (around $300 million) to Bitfinex, generating a total profit of $72.83 million.

Beating AI Insight News Brief: AI lab Supersonic Labs has launched Julia-1, a lightweight decision-making model with just 144.3 million parameters. Unlike models that generate long-form text, Julia-1 takes in context, a query, and 2 to 20 candidate answers to directly perform classification, scoring, and yes/no judgments, enabling agents to select tools, handle routing, or determine next steps. Built on the multilingual encoder model mmBERT-small, Julia-1’s model weights are around 550MB, with a focus on low hardware requirements. The developer has completed tests on Apple M4, Intel Core i5-1235U, and Samsung Android tablets. On the M4, the median latency per decision is roughly 33ms; on Samsung tablets running solely on CPU, latency is around 203ms, with a peak process memory usage of about 393MB. In four sets of comparative tests against Jev announced by the developer, Julia-1 took three wins. Its scores hit 73.15% on Typed Decisions (vs Jev’s 72.70%), 94% on AG News (vs 91%), and 86% on sentiment classification (vs 48%). However, on Banking77—a dataset with 72 categories—Julia-1 scored just 64%, significantly lower than Jev’s 87%. The developer noted that when there are too many highly similar candidate categories, the current grouping screening process may incorrectly discard the correct answer early. Julia-1’s total cloud GPU training and experimental costs amounted to just around $104. The model weights, inference code, full evaluation results, and source documentation are all open-source and released under the Apache 2.0 license. The developer is also developing an API, with plans to charge $0.025 per million input tokens; since the model only performs judgments and does not generate long-form text, output tokens are free of charge.

Beating AI Insight News Brief: OpenAI is reportedly developing an always-on AI assistant codenamed "o". User Jake Boggs spotted an unreleased perk on the ChatGPT Pro upgrade page: "o, your always-on assistant". TestingCatalog later found fields including the display name "o" and email_suffix: -o in ChatGPT’s configuration. The Information reports OpenAI is building features directly competing with Grok Bot, and has discussed a personal AI assistant matching Meta’s Muse. The related products may leverage existing Agent technology from ChatGPT and Codex to run tasks continuously in the background, even across multiple days. Earlier, the community also discovered "gpt-6-astra-aeon" in Codex, plus internal instructions describing Aeon as a "persistent background Agent". This has led some to speculate Aeon is the internal codename, while "o" is the public-facing product name. OpenAI has confirmed its DevDay will take place on September 29, with Sam Altman delivering the opening keynote. TestingCatalog speculates "o" could make its debut at the event.

According to GMGN market data, the Solana ecosystem meme coin ZCAT has rallied nearly 40% over the past 24 hours, with its current market capitalization hitting $82.98 million. ZCAT is a cat-themed meme token built on Solana, drawing inspiration from Zcash’s (ZEC) privacy concept and featuring an anonymous cat mascot with a brown paper bag over its head. The token levies a roughly 3% transaction and transfer tax, which is used to purchase bridged Zcash (ZEC) on the Solana blockchain and airdrop it to holders. BlockBeats reminds users that most meme coins lack real use cases, exhibit high price volatility, and investors should exercise caution.

Whale 0xd0A4 withdrew 18.34M $ENA($5.13M) from #Gate, #Bybit, #OKX and #Binance today.

Insight Beating AI Flash News: OpenRouter has launched Jev Router, integrating the type-safe Jev into its model routing system. Before each request is dispatched, Jev evaluates task difficulty, precision requirements, and the adequacy of the current model and inference tier, then decides whether to keep the existing setup, boost inference intensity, or switch to another model. OpenRouter’s original Auto Router previously selected models based on task type: it first identifies the task category, then references seven days of usage data for similar requests to route. Jev Router goes a step further by assessing the specific difficulty of each individual round. For simple tasks, it can lower inference tiers; for complex ones, it increases reasoning effort. If the same model can still handle the task, it avoids switching unnecessarily. A critical practical factor here is context caching: when switching models during a long conversation, cached context typically cannot be reused, forcing the new model to reprocess the entire chat history. Jev Router accounts for this loss, only switching models if the expected benefit of the switch exceeds the switching cost, prioritizing retaining the same model whenever possible. OpenRouter’s internal testing shows that across four agent benchmarks totaling 423 tasks, Jev Router completed 237 tasks, compared to 130 for Auto Router—an ~82% increase in task completion. Additionally, in five other agent benchmarks, its median first-token latency was lower than that of other tested routers. Similar approaches have previously existed in the Jev community. For example, `gholtzap/jev-codex-model-and-effort-router` uses Jev to select models and inference tiers for Codex, then locks the entire thread to reuse caching; `Jev-Auto-Router` attempts to reselect models per call. OpenRouter has now integrated this approach directly into its official routing service.