Reflection finally delivers its work: 501B-parameter model Beam unveiled, still trailing China’s latest leading open-source models.
Beating AI News Flash: Nvidia-backed AI startup Reflection has unveiled its first open-weight model, Beam. Built on a Mixture of Experts (MoE) architecture, Beam has a total of 501 billion parameters, with only 23 billion activated per forward pass, primarily targeting code, reasoning, and agent tasks. According to Reflection’s own benchmarks, Beam’s overall performance is close to GLM-5.2, though it has not yet caught up with the latest leading Chinese open-source models. On the DeepSWE benchmark, Beam scores 44.4, GLM-5.2 scores 44.0, Qwen 3.8-Max scores 51.0, while GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash reach 61.0, 68.0, and 74.2 respectively. Beam’s training scale is also massive: it used 6,144 Nvidia GB300 GPUs for pre-training, processing 23.8 trillion tokens in under four weeks; it then conducted four consecutive weeks of reinforcement learning (RL) training with 10,500 GB300 GPUs, generating over 100 million training rollouts. The company noted that this is one of the largest RL trainings on public record that they are aware of. Beam is currently in the final security testing phase, with only a small number of users granted early access. Reflection plans to release its full weights, technical report, and model card later this month, under the Apache 2.0 license.
1 minutes ago
Arbitrum Joins Paxos’ Global Dollar Network, USDG Officially Launches on the Arbitrum Network
Ethereum Layer 2 network Arbitrum has announced its membership in the Paxos-led stablecoin alliance Global Dollar Network (GDN), with its issued US dollar stablecoin USDG now officially live on the Arbitrum network. USDG has integrated with DeFi protocols including Morpho, GMX, Fluid, and Maple; Kraken will offer fiat-crypto on/offramps, while platforms like Uniswap are also planning subsequent integrations. Unlike traditional stablecoin models, GDN allocates a portion of reserve asset returns to partners driving USDG adoption, enabling Arbitrum and its ecosystem projects to share the economic benefits from the stablecoin’s growth. Additionally, ArbitrumDAO has received a governance proposal recommending that USDG growth be prioritized as a strategic focus, adding 100 million ARB to the DRIP incentive program, and utilizing treasury assets to support USDG liquidity. The proposal has not yet been approved.
1 minutes ago
Reka launches the 19B Rho-1: a single model capable of processing text, images, video, and robotic actions simultaneously.
Beating AI Express (from Dongcha) reports: U.S. AI startup Reka, founded by former researchers from DeepMind, Google Brain and Meta FAIR, has launched its 19-billion-parameter model Rho-1. The model can process text, images and video in a single conversation, and also output robotic actions. For example, it can first generate a lighthouse image, animate the lighthouse, modify the weather in a video, and then ask about the differences between two videos—all generated content is retained in the context, so there is no need to use another separate model for processing. It also supports receiving instructions mid-video generation: when a video is being generated, users can temporarily request changes to the weather, movement direction or other elements, and subsequent frames will be adjusted accordingly. The base version generates content at approximately 0.79 times real-time speed; the distilled version reduces generation steps from 99 to 8, and Reka measures that it can produce a 5.3-second video in about 1 second. Robotic control is integrated into the same model: Rho-1 predicts what the camera will see next while outputting robotic action signals, eliminating the need for a separate planning model. Currently, official demonstrations are still based on LIBERO simulation tasks, with no physical robot test results released. Similar unified models have emerged this year, such as NVIDIA Cosmos 3, which can also unify the processing of text, images, video, audio and actions. Rho-1 focuses more on continuous interaction, highlighting repeated generation, modification and reasoning within the same context. It was trained from scratch using 320 H100 GPUs over approximately 3 months. The current video resolution is 672×384; long videos may still have frame structure drift, and video localization and partial editing remain unstable. The model is currently a research preview, with weights and APIs not yet open.
1 minutes ago
He Yi: Binance’s Biggest Challenge Is Talent, We Are Eager to Recruit Top Candidates
Binance co-founder He Yi stated during a Chinese AMA session at the Binance Smart Product Launch event that the exchange’s homepage is now fully customizable, allowing users to retain only what’s useful to them. As personalized AI continues to iterate, the mobile homepage could eventually look completely different from others’. Next, Binance hopes to enable users to handle many tasks through communication. Binance Skills already has on-chain capabilities, though the current user experience isn’t very smooth. Moving forward, Binance will integrate these structured, framework-based features into AI Pro. The biggest challenge Binance faces is talent: while not all candidates will be hired (some may fail interviews), the exchange is indeed eager to recruit top talent.
1 minutes ago
Pre-Market News Roundup for US Stocks: AI Duo Accelerate Fundraising, Storage Chip Supply Tightens
Below are key pre-market news items for US stocks on Tuesday:
1. Bridgewater Associates founder Ray Dalio warned that the U.S. could face a debt crisis within three years.
2. AMD’s CEO noted that supply of memory chips remains tight, with a significant supply increase expected by 2027.
3. A Wall Street consortium has launched a record $60 billion AI debt financing deal, designed to fund chip leasing for Anthropic.
4. OpenAI is in discussions with an Emirati fund and BlackRock for $30 billion in financing, with a pre-money valuation of roughly $1.4 trillion.
5. U.S. spot Bitcoin ETFs recorded a net outflow of $89.8 million yesterday, while spot Ethereum ETFs saw a net outflow of $18.9 million.
1 minutes ago