Flash

The news, the context, and what changes.

RSS
Flash

SpaceXSI Launches End-to-End Speech-to-Speech Model Grok Voice 3.0

SpaceXSI has released Grok Voice 3.0, claiming it is the world's most intelligent speech-to-speech model with latency under one second. The model ranks first on the Artificial Analysis comprehensive leaderboard, achieving a 97% score in full-duplex benchmarks and 71.6% in τ-Voice. Its performance is comparable to GPT-Live-1 but with 40% lower latency. Currently, the model supports millions of daily conversations in Tesla vehicles and handles 20,000 calls per day at Starlink customer service cent

88Z AI editorial1 min read
Flash

Kunlun Wanwei Open-Sources Matrix-Game 4.0, a Consumer GPU Video Large Model Game

Kunlun Wanwei has open-sourced Matrix-Game 4.0, the world's first video large model game capable of real-time operation on a single consumer-grade GPU. Running on an RTX 5090, it achieves 480p resolution at 60 FPS with end-to-end latency under 150ms and peak VRAM usage of 18.5 GiB. Its architecture involves a game engine maintaining world state, a Coding Agent generating or rewriting rules, and a video model handling real-time rendering.

88Z AI editorial1 min read
Flash

StepFun Officially Releases Step 5 Preview Model Weights

StepFun has officially released the weights for its Step 5 Preview model, which are now available on Hugging Face and ModelScope. Utilizing a sparse MoE architecture, the model features 600 billion total parameters with only 27 billion activated. It supports a context window of one million tokens and natively handles both text and visual inputs. The model demonstrates strong performance in long-horizon agent tasks such as AI coding and finance, ranking second globally in 第三方模型接口 call volume.

88Z AI editorial1 min read
Flash

Tencent Launches Personal Stock Analysis Agent 'Rena Lancai'

Tencent has launched 'Rena Lancai', a personal stock analysis Agent for ordinary investors, accessible via web, WeChat Mini Programs, and WeChat Assistant. The tool automatically links daily quick reads, market briefs, and watchlist updates. It supports importing holdings, viewing individual stock dynamics, and providing reasons and verification standards for bullish, bearish, or neutral judgments, allowing users to take notes and ask follow-up questions.

88Z AI editorial1 min read
Flash

Google Gemini 4 Argon Officially Open to Global Developers and Enterprise Users

Google's new flagship model, Gemini 4 Argon, is now available to global developers and enterprise users. Access is provided through the Gemini API, paid tiers on the Gemini platform, Google Antigravity Ultra membership, or Ultra and Pro memberships for the Gemini app and search. The model has undergone limited testing and extensive internal real-world validation, focusing on complex multi-step tasks in programming, scientific research, law, and finance.

88Z AI editorial1 min read
Flash

ModelBest and OpenBMB Open-Source Edge Multimodal Model MiniCPM-V 4.7

ModelBest and OpenBMB have jointly released the open-source edge multimodal model MiniCPM-V 4.7, available in 35B-A3B and 1B versions. The 35B-A3B is the world's first open-source MoE multimodal large model capable of lossless, smooth operation on edge devices, requiring only about 3GB of memory and performing well in various benchmarks. The model supports multiple chips and its code is fully open-sourced.

88Z AI editorial1 min read
Flash

Alibaba Qwen Office Launches E-commerce Version for All Merchants

Alibaba's Qwen Office introduces an e-commerce version, an AI business assistant designed for merchants. It connects multiple platforms via a unified entry point, simplifying login, report downloading, and verification processes, while supporting global business overview, trend capturing, ROI optimization, and bulk content generation.

88Z AI editorial1 min read
Flash

Anthropic Launches Git Alternative Scion

Anthropic has released Scion, a self-developed version control system designed to replace Git, which has been in use for two decades. Addressing synchronization bottlenecks and massive traffic pressures caused by the surge in Claude-assisted coding, Scion adopts an always-online centralized architecture. It significantly optimizes code submission and retrieval efficiency through virtual checkouts and streaming pull technologies.

88Z AI editorial1 min read
Flash

Alibaba Qwen Launches Online Reinforcement Learning Framework QwenGyre

Alibaba has launched QwenGyre, an online reinforcement learning framework designed for ultra-long-horizon agent tasks. The framework reduces GPU idling through elastic scheduling and controls trajectory processing to lower the cost of forked training. Tests on the Qwen3.8-2.4T model show a significant increase in NL2RepoBench pass rates for single rollouts of approximately 700k tokens. Under the same GPU budget, it achieves faster acceleration and greater training time savings compared to Async,

88Z AI editorial1 min read
Flash

Meta Launches Frontier Model Series Muse Wonder

Meta has released the new Muse Wonder model and its enhanced version, Wonder Full. These models perform close to Fable 5.1, Opus 5.5, and GPT-6 Astra in various benchmarks. Focusing on active memory and sustained attention, the series aims to realize the vision of personal superintelligence, with specific optimizations for cross-conversation memory, long-process tasks, and coding performance. Zuckerberg emphasized accelerating AI investment and advocating for industry self-regulation to ensure s

88Z AI editorial1 min read
Flash

Zhipu Launches Legal-Specific Large Model GLM-Legal

Zhipu has released GLM-Legal, a legal-specific large model built on the GLM-5.3-Flash base. It covers the entire workflow from contract review to document generation. The model achieved a 19.2% pass rate in Harvey LAB tests, matching leading closed-source models, and supports API and AutoClaw desktop access with a focus on low cost and traceable evidence.

88Z AI editorial1 min read
Flash

Character.AI Files for Bankruptcy, Plans to Sell Partial Assets to Disney for $20 Million

Character.AI has filed for Chapter 11 bankruptcy protection in the US and agreed to sell certain intellectual property assets to Disney for $20 million. Disney serves as the stalking horse bidder with its offer setting the floor price, allowing others to bid higher. The deal is expected to close between late November and early December. Operations continue during the sale process, but the platform will gradually shut down if no buyer is found.

88Z AI editorial1 min read