Gate News, March 25 — Cursor released the Composer 2 technical report, unveiling the complete training scheme for the first time. The base model Kimi K2.5 uses MoE architecture, with a total of 1.04 trillion parameters and 32 billion activated parameters. The training consists of two stages: first, continued pretraining on code data to enhance encoding knowledge; then, large-scale reinforcement learning to improve end-to-end coding capabilities. The RL environment fully simulates real Cursor usage scenarios, including file editing, terminal operations, code search, and tool calls, allowing the model to learn under conditions close to production. The report also details the construction of Cursor’s self-developed benchmark, CursorBench: tasks are collected from actual engineering team coding sessions rather than artificially created. Kimi K2.5 scored 36.0 on this benchmark, and after two-stage training, Composer 2 reached 61.3, a 70% improvement. Cursor states that its inference cost is significantly lower than some cutting-edge large model APIs, achieving Pareto optimality between accuracy and cost.
Disclaimer: The information on this page may come from third parties and does not represent the views or opinions of Gate. The content displayed on this page is for reference only and does not constitute any financial, investment, or legal advice. Gate does not guarantee the accuracy or completeness of the information and shall not be liable for any losses arising from the use of this information. Virtual asset investments carry high risks and are subject to significant price volatility. You may lose all of your invested principal. Please fully understand the relevant risks and make prudent decisions based on your own financial situation and risk tolerance. For details, please refer to
Disclaimer.
Related Articles
Hyperliquid founder Jeff Yan: started with $10k at age 27 to achieve financial freedom, then later founded a decentralized exchange
Hyperliquid founder Jeff Yan reviewed his experiences before creating Hyperliquid in an interview. He once ran a large anonymous trading team, Chameleon Trading, and used $10k to achieve significant growth. Due to dissatisfaction with the state of the crypto industry and the FTX event, Jeff decided to shut down Chameleon Trading and focus on building the decentralized exchange Hyperliquid.
GateNews1h ago
Aptos token overhaul: staking yield slashed in half, Gas fees up 10x as it moves toward deflation
Aptos Foundation updated its tokenomics on April 14. The main reforms include lowering the staking annualized rate to 2.6%, increasing Gas fees by 10x, setting a supply cap of 2.1 billion tokens, and permanently locking 210 million APT. This adjustment is intended to shift the supply model from a “subsidy-driven” approach to a “performance-driven” one, in order to achieve a deflationary target and encourage long-term participation.
MarketWhisper1h ago
Hyperliquid founder: Last year’s biggest challenge was server latency; the team rewrote the system from the ground up.
Hyperliquid founder Jeff Yan said that the biggest challenge in 2025 will be API server latency. Last summer, as Bitcoin’s price surged and trading volume spiked, the system experienced delays: order responses went from instantaneous to 3 seconds, hurting trading efficiency. After the team’s efforts, they rebuilt the server system and ultimately overcame the issue.
GateNews1h ago
Pi Network 18 million users complete KYC, and 26.5 million PI are issued to verifiers
Pi Network distributes 26.5 million PI tokens to more than 1 million community members who have completed KYC verification, to accelerate user identity verification and strengthen the decentralized model. 18 million already-verified users have laid the groundwork for the ecosystem’s development, but the key lies in how to convert these users into active participants. With the protocol upgrade, Pi Network is moving toward the open network phase, and smart contract support signals richer application scenarios.
MarketWhisper1h ago
twin3.ai launches a Web 4.0 soul contract testnet, enabling on-chain digital identities with a 256-dimensional SBT
Twin3 launched digital identity infrastructure for the Agent economy on April 14 and rolled it out on the testnet. The protocol uses soulbound tokens to store users’ identities and experiences, building digital identities, supporting personal AI agents, and driving automated economic collaboration.
GateNews2h ago
Scroll proposes revoking the UN Security Council; multisig addresses take over three major agreement contracts
Scroll Foundation proposed on April 14 to dissolve the current Security Council and transfer administrative control to Scroll Admin’s multisig address because the Security Council’s operating costs do not match actual usage. The transition period will be ten days, and it will rely on member support. The new structure will continue to ensure the protocol’s security standards, and this does not mean compromising on security. The role of DAO contributors will also be scaled back to align with current priorities and operational needs.
MarketWhisper3h ago
