Ant Group's Ling team released Ling-3.2-flash, a natively multimodal model built on a 560B-total / ~25B-active MoE architecture with a 1M context window and image, text and video input. In vendor-run tests it beats Claude Opus 5 on OmniDocBench1.5, ZEROBench and Vision2Web, and enters the DesignArena image-to-frontend top ten at 1238 Elo. API pricing is $0.30 per million input tokens and $0.90 per million output tokens, compatible with OpenAI and Anthropic protocols, and free for the first two weeks.
Long Context Window
长上下文窗口
The maximum number of tokens a model can process in one pass, affecting memory capacity.
Curated Intelligence
Ant Group's open-weight model Ling 3.1 Flash (version 1019) is available for free on the coding agent platform Command Code for four days. The model features 560B total parameters with 25B active, supports a 1M token context window, and targets advanced reasoning and agentic workflows.
Google has released Gemini 4 Carbon, a new frontier model that excels in cybersecurity benchmarks with an 88.2% vulnerability detection rate. It demonstrates capability to handle multi-day engineering tasks on large codebases and significantly outperforms competitors in long-context reasoning. Access is currently limited to select cyber defenders via the Fairwind Program, with introductory pricing at $2/$10 per million tokens.
Liquid AI has released the LFM3 model family based on a new tri-mixed layer architecture, featuring text, vision, audio, and omni-decision variants. The flagship LFM-1.7B outperforms the previous 2.6B model with 35% fewer parameters while supporting a 256K context window, marking a significant leap in small-model efficiency and capability.
Google's Gemini 4 Carbon has secured the #1 spot on the Artificial Analysis Intelligence Index v5. It demonstrates leading performance in complex reasoning and scientific tasks but trails Claude Opus 5.5 in coding agents. Currently available to select users, it maintains standard pricing parity with Gemini 4 Argon while offering a launch discount.
Mirendil has released its first model, Fern 0.1, designed to handle a 10 million token context window (approx. 500k lines of code). While it achieves perfect single-needle retrieval, it lags significantly behind Claude Opus 5.5 in multi-hop reasoning and terminal benchmarks, currently available only via early access.
Baidu has open-sourced the next-generation lightweight document parsing model, ERNIE Unlimited OCR 2. With 3B total parameters but only 500M active during inference, it ranks first on major benchmarks like MarathonDocBench and MPDocBench. The key breakthrough is its ability to parse ultra-long documents continuously without splitting, demonstrated by processing thousands of PDF pages in a single pass while maintaining context, significantly lowering the barrier and cost for long-document AI processing.
Moonshot AI released Kimi K3.1 Pro, which jumped to #8 in the Code Arena: WebDev leaderboard, matching Claude Opus-5 (Max) scores at 70% lower cost. The model features a 2.8T total/104B active parameter MoE architecture with 1M-token context and will soon open-source weights, potentially overtaking DeepSeek-v4.1-Flash as the top open coding model.
Meta and Apple have formed a rare technical partnership, successfully running the open-weight model Muse Spark Open locally on two M5 Ultra Mac Studios equipped with 512GB unified memory via the MLX framework. With 816 billion total parameters and a 1 million token context window, this demonstration marks a significant breakthrough in on-device AI capabilities, despite high hardware costs and pending availability.
Meta has open-sourced Muse Spark Open, an 816-billion parameter multimodal model claimed to be the strongest open-weight model from the US. It supports a 1-million token context and native audio/video processing, optimized for complex coding and long-horizon reasoning. The key breakthrough is its accessibility: it can run locally on just two Apple Mac Studios, eliminating the need for cloud infrastructure or API costs.