Amid Ethereum's rally, on-chain short liquidations exceeded $1.1 billion over the past 24 hours.

Beating AI Express News: Prime Intellect has launched an open-source model inference service called Prime Inference. Users can call models directly on demand; for tasks requiring long-term stable operation, they can reserve a portion of inference computing power in advance. The service’s interface is compatible with OpenAI API, and the first publicly deployed model is GLM-5.3. The system was previously used internally by Prime Intellect. The company states it processes nearly 1 trillion tokens daily, for use cases including reinforcement learning, synthetic data, model evaluation, and coding agents. Prime Intellect has also been running large-scale deployments for clients since January this year. Prime Inference focuses on optimizing long-context workloads for agents. It splits prompt processing and response generation across separate GPUs, cutting p90 inter-token latency by nearly 40% in tests. Prime Intellect also compressed KV cache, boosting the number of tokens each decoder can cache from 1.09 million to 1.63 million under the same GPU memory – an improvement of roughly 50%. Prime Intellect previously offered training infrastructure such as reinforcement learning, evaluation, and sandboxes. With the addition of the inference service, it can integrate model training and actual deployment into a single platform, while data generated from real-world operations can also be used for subsequent training.

According to TradingBeats monitoring, the Loracle address (starting with 0xefe4) currently holds approximately 15,800 SNDK short positions with 10x leverage. The average entry price of the position is around $1,485.67, the current position value is about $27.2 million, with an unrealized loss of roughly $3.7 million, and the position return is approximately -135%. The address continuously added to its SNDK short positions in mid-August, with the short size once rising to around 17,200 units. As SNDK surged in late September, the intraday unrealized loss of its short positions once expanded to about $6.7 million, then narrowed somewhat as the price pulled back. According to BIT (bit.com) market data, as of the latest close, SNDK is trading at around $1,717, up roughly 15.6% from the address's average entry price.

Beating AI News Flash: German AI firm Aleph Alpha has released and open-sourced its English-German bilingual model Kolibri. The model uses a Mixture of Experts (MoE) architecture, with a total of 78.1 billion parameters, though only around 3.46 billion parameters are activated per token. Its weights have been uploaded to Hugging Face under the Apache 2.0 license. Kolibri supports tool calling and adjustable inference modes, with a maximum context window expandable to roughly 1 million tokens. The company recommends limiting daily deployments to within 262,000 tokens, as that is also the maximum context length the model was trained on. In Aleph Alpha’s official evaluations, Kolibri scored 96.9% on AIME 2025 and 85.9% on LiveCodeBench v6. Its combined English and German performance is slightly higher than Qwen3.5 35B-A3B (which also has around 3 billion activated parameters) but remains lower than the fully activated Qwen3.8 (27B).

Binance founder CZ responded on social media to the assertion that Binance’s SAFU fund once bought $1 billion worth of Bitcoin at an average price of $66,666.66, with an unrealized profit of $270 million, noting that it might have been purely luck.

Crypto trader Ansem wrote that constantly monitoring the market and judging price trends based on 15-minute K-lines tends to lead to overtrading. Investors should manage separate accounts for spot trading, perpetual contracts, and high-risk on-chain transactions, and avoid attempting to catch every local top and bottom, as this can erode long-term investment returns. Rather than trading frequently, study past bull markets to observe how long uptrends typically persist. Altcoin rally cycles are usually faster, often outperforming the market for 4 to 6 months before new market leaders emerge—especially after a market cap increase of over 10x. However, this cycle may have a few exceptions: some altcoins will see significant revenue growth alongside price gains, so their fundamentals may improve in parallel. In prior cycles, altcoins mostly peaked due to shifting market attention and fading momentum, but this time, if projects post substantial improvements in metrics like revenue, institutional funds may continue buying actively, and investors need to adjust their existing judgments promptly based on new information. In the last cycle, Bitcoin bottomed in January 2023 and peaked in October 2025, with its uptrend lasting roughly 33 months. If this cycle bottomed in July, it is only in its 4th month currently. He notes that longer-term bull markets will see multiple "mini bull runs" led by specific altcoins, and investors need to identify these period-specific strong assets, reinvest profits into high-performing targets, and gradually convert short-term winners into long-term investments.

According to official announcements, HyperLink has secured $2.5 million in funding, with investors including Alliance, North Island Ventures, Reverie, Node Capital, Breed VC, and SmartestMoney. Positioned as a prime broker for Hyperliquid, HyperLink focuses on low fees, private order flow, and optimal execution. It reduces trading costs via volume aggregation and staking to unlock VIP rates, while masking user wallet addresses—markets only see HyperLink’s transaction flow, and the platform can access Hyperliquid’s full liquidity through the same API. HyperLink stated that its platform routed a cumulative $254 million in trading volume over the past month, accounting for roughly 0.1% of Hyperliquid’s total volume. Currently in an invite-only phase, it already makes up 1% of Hyperliquid Builder’s trading volume. The company’s next goal is to increase its share of Hyperliquid’s total trading volume to 10%.