Tag with "AI Models"

Xiaomi has rolled out a subscription service for its MiMo large language model, offering four tiers from 39 to 659 yuan monthly. The move simplifies AI access for developers while showcasing Xiaomi's growing capabilities in the competitive AI space. With MiMo-V2-Pro ranking fifth globally, this could mark a shift from tech demos to practical applications in China's AI landscape.

Xiaomi has rolled out its first paid subscription plans for the MiMo large language model, marking a shift from free beta testing to commercial service. With four pricing tiers starting at just 39 yuan per month, developers and AI enthusiasts can now access Xiaomi's professional-grade AI tools. The plans include full access to three core models covering text, voice, and multimodal capabilities - a strategic move as Chinese tech firms seek sustainable AI business models.

OpenRouter has quietly introduced two enigmatic AI models—Hunter Alpha and Healer Alpha—that are sparking intense speculation. Hunter Alpha boasts a staggering trillion parameters and specializes in complex reasoning, while Healer Alpha shines in multimodal understanding. Both currently operate anonymously and offer free access, leading to intriguing theories about their origins.

NVIDIA is making its boldest move yet beyond chips, pledging $26 billion to develop open AI models. This strategic shift aims to transform the company from hardware provider to full-stack AI powerhouse. Their Nemotron 3 Super model already shows promise, outperforming rivals in benchmarks. The investment signals NVIDIA's ambition to shape the future of AI development while strengthening its ecosystem.

DeepSeek V4 Lite, a surprisingly powerful AI model with just 200 billion parameters, is turning heads in the tech community. Originally launched in February with strong long-context processing capabilities, recent updates have dramatically improved its performance. Developers report it now rivals top international models like Anthropic Claude 3.5 Sonnet in logic, programming, and aesthetics. This unexpected leap forward has sparked excitement about what its full version might achieve.

Alibaba's Qwen team has unveiled significant upgrades to its open-source AI model family. The expansion introduces three new models targeting different performance needs, from complex reasoning tasks to lightweight applications. Alongside these releases, Alibaba Cloud launched Qwen3.5-Flash API, a managed service supporting up to 1 million tokens context length.

A mysterious new AI model called Pony Alpha has appeared on OpenRouter, offering surprisingly powerful capabilities—completely free. Early tests show it rivals top-tier models like Claude Opus in coding and reasoning tasks. While its origins remain unclear, many suspect it might be Zhipu AI's upcoming GLM-5 model testing the waters. The sudden appearance raises questions about why such advanced tech is being given away—and what the developers might be learning from user interactions.

ZenMux's latest update brings Claude Opus 4.6 to its free tier for two weeks, giving users unprecedented access to cutting-edge AI capabilities. This Anthropic-powered model boasts impressive features like million-token memory and multi-agent collaboration, outperforming competitors in coding and analysis tasks. While the free version has some limitations, it's a golden opportunity for developers and curious minds to test drive premium AI without opening their wallets.

Xiaomi has introduced payment options for its MiMo-V2-Flash AI model while offering free quotas to users. The tech giant revealed pricing details for both domestic and international markets, with the model boasting impressive 309 billion parameters. While requiring real-name verification in China, overseas users enjoy simpler payment methods through Apple Pay and credit cards.

Xiaomi is giving AI enthusiasts more time to play with its powerful MiMo-V2-Flash model. Originally set to expire at the end of December 2025, the free trial period has been extended by 20 days until January 20, 2026. This open-source model packs serious power with 309 billion parameters and shines in reasoning and coding tasks. While keeping access free for now, Xiaomi is preparing to roll out payment options soon.

French AI startup Mistral has unveiled its Mistral3 series, offering powerful open-source models that outperform competitors while running efficiently on modest hardware. The lineup includes compact yet capable models ranging from 3B to 14B parameters, all featuring an impressive 128K context window. What really turns heads? The flagship model delivers these results using just a single A100 GPU while maintaining commercial-friendly Apache licensing.

MiniMax has launched its M2 inference model, boasting 230 billion parameters and speeds of 100 tokens per second. Designed specifically for smart agents, the model emphasizes efficiency and real-time interaction capabilities while reducing context window size for practical deployment.

Ant Group's BaiLing large model team has open-sourced Ring-flash-linear-2.0-128K, a cutting-edge AI model combining hybrid attention and MoE architecture. The innovation achieves state-of-the-art performance in long-text programming with only 6.1B activated parameters, offering significant efficiency gains.

Baidu's PaddleOCR-VL model has topped the global OCR rankings with a score of 92.56 in OmniDocBench V1.5, surpassing competitors like DeepSeek-OCR. It has also led Huggingface's trending list for five consecutive days, supporting 109 languages and advanced document parsing.

Alibaba's Qwen team has launched two lightweight versions of its Qwen3-VL visual language model series - 4B and 8B parameter variants. Despite their smaller size, these models maintain strong performance across multiple benchmarks, sometimes approaching the capabilities of their larger 72B predecessor. The release marks significant progress in making powerful AI models more accessible for edge and consumer devices.

LiquidAI has launched the 'Liquid Nanos' series, a collection of lightweight AI models optimized for edge computing devices like Raspberry Pi. The models, available in 350M and 1.2B parameter versions, support five key applications: translation, extraction, RAG, tool calling, and mathematical reasoning. Twelve task-specific models are now live on Hugging Face.

Baidu's newly open-sourced ERNIE-4.5-21B-A3B-Thinking model has claimed the top spot in Hugging Face's text generation rankings. The lightweight MoE architecture model demonstrates exceptional reasoning capabilities while maintaining high parameter efficiency, marking a significant advancement in China's AI ecosystem.

Anthropic confirmed temporary quality declines in its Claude AI models due to technical failures but has since resolved them. The issues affected Claude Sonnet4 and Haiku3.5 requests between August and September 2025. The company emphasized these were not cost-cutting measures and thanked user feedback for aiding swift resolution.

Silicon-Based Flow has partnered with Ant Group to launch Ling-mini-2.0, an open-source AI model featuring MoE architecture. With 16B total parameters but only 1.4B activated per token, it achieves remarkable speed while maintaining high performance across coding, math, and reasoning tasks.

Microsoft has launched its first self-developed AI models, MAI-Voice-1 and MAI-1-preview, marking a significant step in its competition with OpenAI. MAI-Voice-1 excels in rapid audio generation, while MAI-1-preview targets instruction-based tasks. Both models are set to enhance Microsoft's consumer AI offerings.