Jiang Zhuoer: Bitcoin may fall below $80,000, with the $77,000 support level expected to hold

Insight Beating AI News Flash: Zhipu AI has launched GLM-Legal, a legal-specific model further trained based on GLM-5.3-Flash. In the Harvey LAB evaluation covering 24 types of legal services, its task pass rate rose from 6.7% to 19.2%—nearly triple the original level, outperforming models including Kimi K3 and Claude Fable 5, and ranking 7th on the Vals AI leaderboard. Zhipu demonstrated several practical application cases: when reviewing software service contracts, GLM-Legal suggested shortening the termination notice period from 60 days to 30 days, and required the service provider to provide up to 90 days of handover support. In another software project dispute involving 40 documents, it found that the company reported 60% project completion to clients while internal records showed only around 41%, uncovered emails from the CEO asking to hide risks and continue charging, and analyzed potential legal liabilities based on contract terms. According to Vals AI's independent evaluation, GLM-Legal costs only $0.15 per test, while the comparable Gemini 4 Argon costs $10.55—an approximately 70x cost difference. However, GLM-Legal's average processing time reaches 72 minutes, much slower than Gemini 4 Argon's 44 minutes. Zhipu stated that the model was trained on specialized legal data and underwent reinforcement learning in tens of thousands of simulated legal environments to enhance its ability to complete complex tasks. Currently, API trial access for GLM-Legal requires application.

Amid the sell-off of memory chip stocks, Elon Musk posted remarks that once China masters the manufacturing technology of a product, its capacity to scale production often exceeds that of the rest of the world combined—a trend already observed in the lithium battery and solar sectors. He noted that the memory industry may be an exception, but roughly three years ago, memory was still a low-margin commodity. LP5 and LP6 follow universal standards, and China appears to be rapidly mastering the relevant technologies. However, next-generation HBM architectures, as well as technologies such as placing an SRAM layer beneath logic circuits, 3D-stacking memory on top, and resolving logic circuit heat dissipation simultaneously, will be far more challenging to master. Musk also said SpaceX is a rare counterexample, with its total mass of payloads sent to orbit roughly an order of magnitude higher than that of the rest of the world combined. However, of the total mass of payloads sent to orbit excluding SpaceX, China accounts for the majority. Previous reports noted that Changxin Technology announced key progress in the development of the next-generation DRAM 4F2 architecture, and is expected to launch DDR5 RDIMM server memory products integrating this architecture and technologies like hybrid bonding by the end of the year. Compared with the mainstream 6F2 architecture, the 4F2 theoretically reduces cell area by about one-third, boosts storage density by roughly 50% at the same wafer area and process node, stoking market concerns over future competition and supply expansion.

According to on-chain analyst Ai Yi (@ai_9684xtpa), a certain VVV smart money address has reduced its position again after a half-month hiatus. The address deposited 100,908 VVV tokens into Coinbase 3 hours ago, worth approximately $2.34 million. Over the past four months, it has accumulated total deposits of 562,000 VVV tokens at an average price of $21.26, for a total value of $11.97 million, with estimated per-token profits exceeding $9.84 million.

According to on-chain analyst Yu Jin Monitoring, as of press time, a whale or institutional entity has sold a total of 837.8 BTC and purchased 27,500 ETH worth approximately $69.68 million, with an average purchase price of $2,534 for ETH. The entity’s transaction flow: It transferred on-chain BTC withdrawn from HTX a year and a half ago to Hyperliquid, sold the BTC for USDC; then withdrew the USDC to Arbitrum, bridged it to Ethereum, and finally bought ETH on-chain.

Multiple sources familiar with the matter told Caijing Magazine that large language model startup DeepSeek will soon close its latest funding round, with the total capital raised rising from the previously planned 50 billion yuan to around 100 billion yuan. This marks DeepSeek’s second financing round, where the planned investment amount has doubled in less than half a month. New and existing shareholders will each account for roughly half of the total funding. Despite the doubling in financing scale, DeepSeek’s pre-money valuation remains unchanged at approximately 500 billion yuan. New investors in this round include Shixi Capital, SMIC Pi Capital, CPE Source Capital, Legend Capital, and Geely, among others, with industrial investors forming the majority of the new backers. (Caijing Magazine)

Beating AI News: The long-teased X Subscription has officially launched. The new offering bundles X Premium, Grok, Grok Bot, and Cursor into a single subscription, with four paid tiers priced at $8, $30, $100, and $200 per month, plus a free tier. The most notable tier is the $30 monthly Premium plan. Per official details, this package combines benefits from the previously separate X Premium+, SuperGrok, and Cursor Pro—totaling $90 monthly in individual costs—plus access to Grok Bot, Cursor’s cloud Agent, and code review features. The two higher tiers primarily increase usage allowances, set at 15x and 50x the $8 base plan’s limits, respectively. However, the new billing structure differs from prior plans: Grok, Grok Bot, and Cursor share a single weekly-refreshed allowance, which applies mainly to Grok’s native models. The company states that the new plan’s Grok model programming allowance is at least double that of the old Cursor Pro plan. That said, accessing third-party models like Claude and GPT via Cursor will still require additional payments based on API rates. Existing X Premium, SuperGrok, and Cursor members can either retain their current plans or choose to migrate. Eligible old subscriptions will receive pro-rated refunds for remaining time, but account merges are irreversible—users cannot revert to their original plans after switching.