Metaplanet transferred another 1,350 BTC to Coinbase Prime an hour ago, valued at approximately $108 million.

According to Bitget market data, South Korea’s KOSPI index closed down 123.48 points, or 1.79%, at 6788.89 points on Friday, August 28. SK Hynix fell 4.45%, and Samsung Electronics dropped 3.38%.

Morgan Stanley has adjusted ratings for multiple European industries, upgrading the rating of the European aerospace sector from "Equal Weight" to "Overweight".

Beating AI News: Tencent has published results from a blind engineering test for its Hy4 model. 163 internal Tencent experts evaluated different models across 203 real-world engineering tasks, with Hy4 posting an average score of 2.99/4, outperforming GLM-5.3 (2.92/4) and Kimi K3 (2.94/4). Specifically, Hy4’s win rate against GLM-5.3 stands at 46.8%, with a 12.8% draw rate and 40.4% loss rate; versus Kimi K3, its win rate is 51.2%, draw rate 7.9%, loss rate 40.9%. The test tasks are actual internal Tencent engineering projects, not traditional fixed benchmarks. However, Hy4 does not lead across all public benchmarks. Per Tencent’s released evaluation data, it has mixed performance against models like GLM-5.3 and Kimi K3, and still lags behind GLM-5.3 in code and cybersecurity tests including DeepSWE and CyberGym. Price is Hy4’s most prominent advantage: it costs 6 yuan per million input tokens and 18 yuan per million output tokens—25% cheaper than GLM-5.3 for input and roughly 36% cheaper for output; compared to Kimi K3, it is 70% cheaper for input and 82% cheaper for output. Its cache hit cost is just 0.3 yuan, 85% lower than GLM-5.3 and Kimi K3’s 2 yuan.

According to a report by Seoul Economic Daily, Samsung Electronics is developing an 8-layer HBM4E product as required by NVIDIA, for use in NVIDIA’s custom high-bandwidth memory NVHBM. This design cuts the number of stacked layers compared to Samsung’s original plans for 12-layer and 16-layer products. NVIDIA has set a target speed of 17 to 18 Gbps, roughly 20% higher than the 14.4 Gbps of Samsung’s early HBM4E samples. The 8-layer design reduces the difficulty of wafer thinning, precision stacking, and backend packaging, eases yield pressure, and boosts supply capacity. NVHBM is expected to debut in NVIDIA’s Rubin Ultra AI GPU, scheduled for launch next year. Rubin Ultra will expand its GPU interconnect scale from the current 72 units to up to 576 units, using more high-speed GPUs working in tandem to offset the drop in storage capacity per individual GPU. Samsung has the capability to design and manufacture DRAM, logic chips, and base dies, allowing it to offer one-stop services for custom HBM. Industry insiders point out that Samsung has already demonstrated high transmission speeds on HBM4, and as NVIDIA shifts its competitive focus from high stacking to high-speed transmission, Samsung’s related capabilities may draw more attention.

Beating AI News: Tencent has released the Hy4 preview, a new Mixture of Experts (MoE) architecture model with 770 billion total parameters, 49 billion activated parameters per inference, and a context window expanded to 1 million tokens. Compared to Hy3’s 295B total parameters, 21B activated parameters, and 256K context window, this generation’s scale has more than tripled, while the context window is roughly four times larger. This time, Tencent did not solely emphasize general benchmark scores, but specifically enhanced the model’s performance on real-world tasks including coding, office work, gaming, and scientific research. For example, the model can build a Three.js 3D website from scratch, create playable game demos in Unity, and process up to 72 financial documents at once to check for duplicate reimbursements, over-limit expenses, and budget anomalies. Hy4 was also trained and iterated alongside Tencent products such as WorkBuddy and CodeBuddy. A standout feature is that Hy4 has begun participating in its own R&D: Tencent says the model helps optimize training methods, data strategies, evaluation systems, and underlying operators, proposing its own plans, running experiments, and adjusting based on results. The code, logs, and feedback from these experiments are fed into the next R&D cycle, forming an initial self-improvement loop. The Hy4 preview is now integrated with WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub, and OpenRouter. Tencent previously teased in its financial reports that Hy4 would be larger than Hy3, and noted real product feedback as a key source for model training.

Beating AI Express News reports that Cognition, the company behind AI coding agent Devin, has reached an annualized revenue of approximately $900 million. Calculated at its current monthly revenue of around $75 million, this represents a more than threefold increase from the start of this year. The firm is also in a new round of financing, with a potential valuation of roughly $45 billion, and its growth pace is accelerating. In May this year, when Cognition secured its previous financing, it disclosed an annualized revenue of just $492 million and a $26 billion valuation. In only three months, its annualized revenue has climbed by about 83%. However, Cognition is burning cash heavily. The Information states that the company is projected to consume roughly $800 million in cash this year, with $200 million of that burned in the second quarter alone. A major portion of this spending goes toward computing power: Cognition leases a large fleet of NVIDIA servers, used both to operate Devin and train its own models, with the server setup costing hundreds of millions of dollars annually. Currently, Cognition’s enterprise business has a gross profit margin of nearly 50%. Excluding spending on training its in-house models, the company is nearly cash flow neutral. Management forecasts that annualized revenue will exceed $1.5 billion by the end of this year, and reach $4 billion to $5 billion next year.