Gate News message, April 24 — DeepSeek V4 has published results from formal mathematical reasoning evaluations, achieving a perfect score of 120/120 on Putnam-2025, tying with Axiom for first place. In the practical regime using LeanExplore and constrained sampling, V4-Flash-Max scored 81.00 on the

Gate News message, April 23 — CryptoQuant's Bitcoin Bull Market Score Index has rebounded to the neutral level of 50 for the first time since Bitcoin declined from its high of $126,000, signaling a potential shift from bear to neutral market conditions. The index, which synthesizes ten on-chain

Alibaba's Happy Horse-1.0 model has achieved the highest ranking on the Arena Video Edit global benchmark with an Elo score of 1,299, outperforming xAI's Grok Imagine Video and Kling o3 Pro by significant margins.

On March 31, RootData released a crypto market-maker transparency ranking report. Galaxy Digital, Wintermute, and GSR ranked in the top three, and their transparency performance was all above 70%. Most market makers with transparency below 60% did not disclose key data. The scoring takes multiple factors into account; a higher score indicates stronger capabilities and greater stability.

Microsoft announced new features for Copilot Researcher called Critique and Council, combining OpenAI's GPT and Anthropic's Claude to improve the quality of research tasks. Out of 100 complex tasks, the Critique mode scored 57.4, significantly ahead of other models.

ARC Prize Foundation Launches ARC-AGI-3 Benchmark, Requiring AI to Independently Explore and Act in a 64×64 Grid World Without Guidance, with Scores Based on "Action Efficiency." Compared to previous versions, this edition is more rigorous, preventing models from using memory shortcuts. Google Gemini 3.1 Pro Preview Achieves a Top Score of 0.37%.

Cursor released the Composer 2 technical report, introducing a complete training solution for its Kimi K2.5 MoE architecture, including two-stage training and a proprietary benchmark called CursorBench. After training, Composer 2 demonstrated significant performance improvements and outperformed other cutting-edge models in terms of inference costs.

Cursor released the Composer 2 technical report on March 25, revealing the training scheme of the Kimi K2.5 model, which adopts a MoE architecture with parameters reaching 10.4 billion. Training is divided into two stages, using real-world scenario simulation for reinforcement learning. The model ultimately achieved 61.3 points on the CursorBench benchmark, a 70% improvement, with inference costs lower than other large language model APIs.

xAI's Grok 4.20 officially ended beta testing on March 19th and launched four reasoning modes. Grok 4.20 set a record of 78% in hallucination-free rate testing, but its composite intelligence score is 48 points, lower than competitors. API pricing is $2 for input and $6 per million tokens for output, supporting context of up to 2 million tokens.

On March 18, on-chain data analyst Axel released a report indicating that the Bitcoin market has moved out of an overheated state, but selling pressure still exists and no reversal signals have appeared. The MVRV Z-Score indicator has dropped to 0.674, suggesting that the bubble has been deflated; meanwhile, the aSOPR indicator remains below 1.0, indicating that the market is still in a loss-selling phase, with a rebound facing the risk of selling pressure.