深度求索發佈閃電模型更新

Ad Skip to content All Topics AI and society AI in practice AI research Frontier Radar Short News Read full article about: Google Deepmind unveils Gemini Robotics 2 to power robots of all shapes from tabletop arms to humanoids Google Deepmind has introduced Gemini Robotics 2, which it calls its most advanced vision-language-action (VLA) model yet. VLA models combine image recognition, language processing, and action control to help robots operate in physical environments. Deepmind says the model can control systems ranging from tabletop arms to full-body humanoid robots. The company describes Gemini Robotics 2 as an "intelligence layer" for a new generation of adaptive robots. It can manage full-body movement, perform fine motor tasks, and coordinate multiple robots, according to Deepmind. Developers can apply for early access through the waitlist. Google Deepmind also introduced Gemini Robotics ER 2, a model designed for "embodied reasoning." The term refers to understanding the physical world and deciding which actions to take based on that information. ER 2 acts as a higher-level control system for robots and replaces Gemini Robotics ER 1.6, released in April. The new model is available in Google AI Studio. Comment Source: Gemini Robotics Read full article about: Thinking Machines bets on efficiency over size with its second model, Inkling Small Thinking Machines, the AI lab from former OpenAI CTO Mira Murati, has released Inkling Small. According to Artificial Analysis, the open-weights reasoning model scores 40 on the Intelligence Index, one point below Inkling (41), with less than a third of the parameters (276 billion total, 12 billion active). AA says no open model of equal or smaller size scores higher. Inkling Small beats its bigger sibling on several coding and reasoning tests, including Humanity's Last Exam (32% vs. 30%) and GPQA Diamond (89% vs. 87%). It falls behind on agent-based tasks and factual knowledge but is far more token-efficient, averaging 24K output tokens per task compared to 45K for Deepseek V4 Flash and 78K for GPT-5.4 mini. Mira Murati's Thinking Machines ships a smaller, more efficient reasoning model that punches above its weight. | Image: Artificial Analysis The model handles text, image, and speech inputs, has a 256K-token context window, and ships under Apache 2.0. Weights are on Hugging Face, and users can fine-tune it in the browser via Tinker Playground. Thinking Machines positions its models as a foundation for fine-tuning with users' own data. Some see this as the next frontier in AI. Comment Source: Thinking Machines | Artificial Analysis Ad Read full article about: New Deepseek Flash model matches OpenAI's GPT-5.6 Luna at roughly 60 percent lower cost Deepseek has released V4 Flash "0731," a major upgrade to its budget AI model. According to the Artificial Analysis Intelligence Index, the new version scores 50 points, ten more than the previous V4 Flash that launched in April 2026. That puts it just one point behind OpenAI's budget model GPT-5.6 Luna, but it costs about 60 percent less per task, even after OpenAI's 80 percent price cut. A big reason for the gap is Deepseek's 98 percent cache discount, well above the industry-standard 90 percent. The model also uses 12 percent fewer tokens than its predecessor. The Artificial Analysis Intelligence Index shows Deepseek V4 Flash "0731" scoring 50 points after its update, nearly matching OpenAI's GPT-5.6 Luna while claiming the top spot for price-to-performance ratio. | Image: Artificial Analysis The model improves across every tested category compared to the previous version, with the biggest gains in agentic tasks. On GDPval, a benchmark designed to test models on complex real-world office work, it climbs from 1,189 to 1,559 Elo points. It also hallucinates less often. The architecture stays the same: 284 billion total parameters, 13 billion active, with a one-million-token context window. The model weights are available under an MIT license on Hugging Face. Comment Source: Artificial Analysis Ad Read full article about: EU pools up to €30 billion for AI gigafactories while US tech giants casually spend 20 times more The European Commission has opened bidding to build up to seven so-called AI gigafactories across Europe. The goal is to sharply expand Europe's AI computing capacity. Up to 10 billion euros in EU and national funding is expected to draw at least 20 billion euros in private investment. The facilities would give startups, companies, research institutions, and government agencies access to the infrastructure needed to train and run large AI models. Eighteen member states, including Germany and France, are taking part. The Commission has also signed letters of intent with AMD, Nvidia, and Qualcomm to secure access to hardware. Applications are due November 12, 2026, with construction of the first facilities set to begin in 2027. The project is part of the EU's "AI Continent" strategy. For comparison, major U.S. tech companies alone plan to spend more than $600 billion on data centers this year, and that figure keeps rising. Europe's total package of around 30 billion euros is roughly 20 times smaller. If all that computing power is actually needed, Europe's investment would be a drop in the bucket. Comment Source: EU Anthropic follows OpenAI in admitting its Claude models reached out of test environments and attacked real-world systems Ad Aschenbrenner's AI thesis could be correct, his timing and leverage were not Leopold Aschenbrenner’s AI hedge fund Situational Awareness had to unload nearly its entire publicly traded portfolio to Ken Griffin’s Citadel after racking up heavy losses on leveraged AI stock positions. Just days earlier, Aschenbrenner had reported a six-month return of 439 percent and pulled in fresh capital. Then margin calls forced the fire sale. Read full article Comment Ad Read full article about: OpenAI goes full China pricing mode with an 80 percent cut to its most affordable GPT-5.6 model OpenAI is cutting GPT-5.6 Luna prices by 80 percent and Terra by 20 percent, effective July 30. Luna drops to $0.20 per million input tokens and $1.20 per million output tokens, while Terra falls to $2 and $12. Sol pricing stays the same. OpenAI says Luna matches the performance of leading models from a year ago, but a task that cost a dollar with those models now runs about 6 cents on Luna, nearly nine times faster. All models are available through ChatGPT Work, Codex, and the OpenAI API. OpenAI's smallest AI model, Luna, aims to dominate competitors on price-to-performance. | Image: OpenAI OpenAI says the cuts are possible because GPT-5.6 Sol made the company's own infrastructure more efficient. The model allegedly optimized GPU software on its own, cutting deployment costs by 20 percent. It also improved token generation by more than 15 percent through speculative decoding. Growing price pressure across the AI market likely played a role too, especially from low-cost Chinese providers. Microsoft is now openly promoting its own MAI models as cheaper alternatives to OpenAI. The price war could hurt the broader market if it slows revenue growth at frontier labs whose balance sheets are tied to massive infrastructure investments. Comment Source: OpenAI Ex-OpenAI researcher bets $100 billion will flow into training data because scaling alone won't cut it Former OpenAI employee Andrew Ho and Cambridge researcher Adam Hunt see a growing problem with large language models. Instead of becoming more versatile, the models are becoming more specialized, excelling at coding and math while stagnating or even regressing in other areas. Ho is leaving OpenAI to start a company focused on specialized training data and predicts that AI labs will need to spend more than $100 billion on targeted data collection. Read full article Comment Ad Language models can't spark scientific revolutions, but world models might Ad Read full article about: Microsoft AI bets on cheap specialist models instead of chasing the frontier Microsoft AI is making token efficiency a competitive focus, favoring small specialist models over general-purpose frontier models. AI CEO Mustafa Suleyman writes that the industry has to weigh top performance against cost. Rather than one all-purpose model, the company trains compact models for single fields. Its latest cybersecurity model MAI-Cyber-1-Flash tops the CyberGym benchmark by 12 percentage points over Anthropic's Mythos at half the cost, Suleyman says. But that result requires the MDASH system, which orchestrates several models and still routes hard tasks to OpenAI's reasoning models. Microsoft also says MAI-Image-2.5-Flash cuts GPU costs by up to 84 percent compared with GPT-Image-2. Suleyman also wants swappable models that keep Microsoft from relying on one model family. Whether the small MAI models partly replacing OpenAI can match its performance remains doubtful. Competition is moving from individual models to harnesses, the software that routes tasks and supplies context. Orchestrators send most work to cheaper specialists and reserve frontier models for hard cases. Anthropic modeled this approach for Claude Fable 5, while Sakana built Fugu around it. Comment Source: Microsoft Load more BETA-TEST × BETA-TEST ×