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.
Native Multimodality
原生多模态
Models trained from scratch to process text, images, and other modalities within a unified framework.
Curated Intelligence
Moonshot AI has released the open-weight Kimi K3.1 Pro, its most capable 3T-class native multimodal agentic model to date. Featuring a Mixture-of-Experts architecture with 2.8T total parameters and 104B activated parameters, it supports a 1-million-token context window. Post-trained via large-scale joint reinforcement learning, the model delivers substantial improvements in coding, 3D creation, financial research, and factual accuracy.
Meituan's LongCat team released LongCat-2.5, a native multimodal model with 1.6 trillion total parameters and ~48 billion activated per token. The update focuses on enhancing visual understanding and agentic coding, leveraging large-scale long-horizon reinforcement learning to improve end-to-end engineering capabilities for complex codebase development and full-project builds.
Mistral AI has released the weights for Mistral Large 4 (ML4), a 1-trillion parameter native multimodal MoE model. It matches leading open-source models on coding agent benchmarks while reducing average token consumption per task by 47%, significantly optimizing inference costs.
DeepMind has released Genie 4, a new native multimodal world model that synchronously generates visuals, audio, and interactive dialogue. By integrating capabilities from the Gemini family, Genie 4 significantly enhances realism and coherence in real-time interactive simulations, marking a crucial step toward AGI. Users can currently join a waitlist for early access.
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.
Meta has released Muse Spark Open, a native multimodal model with 816B total parameters (42B active) and a 1M-token context window, claiming it as the strongest US open-weight model. Optimized for coding and agents, it uniquely supports local execution on just two Apple M5 Ultra Mac Studios and is available via major cloud APIs and inference engines.
Meta has released Muse Spark Open, claiming it to be the most capable US open-weight model to date. Featuring an 816B total/42B active parameter MoE architecture, a 1M token context window, and native multimodality, it is optimized for long-horizon agentic tasks and complex software engineering, now available for testing on Arena.
HiDream has released HiDream-O1-Video-1.0, a native omnimodal video model ranking #6 on the Artificial Analysis Image-to-Video with Audio leaderboard. The model focuses on physical consistency, generating 1080p videos (5-20 seconds) with natively synchronized audio where duration is determined by scene content rather than fixed presets. It is currently available via the HiHarness API at $5.80 per minute.
Xiaomi announced the deprecation of its open-source LLMs MiMo-V2.5-Pro and MiMo-V2.5 in October 2026, requiring users to migrate to the newly released MiMo-V2.6 series. The V2.6 lineup features native multimodal Pro and Flash models, retaining the previous API pricing structure to deliver comparable intelligence at a fraction of overseas costs.