Tag
Meta 发布 Muse Spark Open，号称迄今最强大的美国开源权重模型，专为编码与智能体任务打造，拥有 816B 总参数（42B 激活）、1M 上下文窗口及文本/图像/视频/音频多模态能力，并支持多种推理引擎与微调框架。
A technical report for a 10M context window model has been published by Voltropy, potentially revolutionizing AI's capacity to handle extremely long texts.
Xing4.0 is an AI model from China Telecom with a 29B/4B Mixture of Experts architecture, trained on Ascend 910C using MindSpore, offering up to 512K context and open-sourced under Apache 2.0.
dots3-note Preview is a multimodal Mixture-of-Experts AI model with 280B total parameters and 16B activated, supporting up to 512K token context, released as an open-weight model on Hugging Face.
Xiaomi has released two state-of-the-art AI models, MiMo-V2.6 Pro RL for maximum capability and MiMo-V2.6 Flash RL for maximum efficiency, both featuring sparse mixture-of-experts architecture, 1M context length, MIT licensing, and native omni-modal support for text, image, video, and audio.
Novita announces support for the DeepSeek-V4.1-Flash AI model on Hugging Face, featuring 552B parameters, native image and text input, a 1M-token context window, and controllable reasoning effort.
Z.ai has officially released GLM-5.3-Flash, a natively multimodal AI model with 320B parameters and a 1M-token context window, previously previewed as Ox Alpha, available via Modular Cloud and running on Chinese AI chips under an MIT License.
An anonymous lab released the Ox Alpha model on OpenRouter, featuring 1M token context, multimodal capabilities, and free access. It was showcased in an agent that autonomously built and verified a Three.js website.
RedNote has released dots3-note preview, the first open-weight model in the dots3 family, featuring 280B parameters with 16B active, multi-modal understanding (text, image, video, audio), 512K context length, and Apache 2.0 license, with strong agent capabilities.
Santiago Valdarrama reports that Kimi K3 is the best open-weight model he has tested, a 2.8T-parameter vision model with tool calling, reasoning, and a 1M context window.
Inkling, a 975B Mixture-of-Experts model (41B active) with 1M context and Apache-2.0 license, introduces a novel audio front end using a 7.9M parameter lookup table instead of a traditional encoder, achieving strong performance on speech tasks. The model was pretrained on 45 trillion tokens of text, images, audio, and video.
Kimi K3, the latest open-weight model from Moonshot, launches with a new architecture, agent swarm capabilities, and focus on long-horizon agent workflows, positioning it as a major contender against other top models.
NVIDIA AI releases a 75B MoE model (9.3B active) compressed from Nemotron-3-Super-120B using the Iterative Puzzle framework, with 1M token context support.
Google quietly made its newest flash model, Gemini 3.5 Flash, free-tier eligible with no credit card, a 1M context window, and 1,500 requests per day, including native multimodal support.
Mimo 2.5 demonstrates fast performance with large context windows using dual RTX Pro 6000 GPUs.
Zhipu AI (zai_org) has open-sourced GLM 5.2, a flagship model for coding and long-horizon agentic tasks with a usable 1M-token context. Modular is a Day Zero launch partner, offering optimized serving on Modular Cloud.
The user reports that the Gemma 4 12B unified audio model stops attending to speech when the system prompt is large (~21k tokens), and asks for workarounds or explanations, noting the issue persists across vLLM, llama.cpp, and LiteRT-LM backends.
NVIDIA released Nemotron Ultra, a hybrid MoE model with 55B/550B parameters and a 1M context window, supporting MTP speculative decoding and available day-0 in transformers.
MiniMax released M3, an open-weights model combining frontier coding, 1M context, and native multimodality, offering comparable performance to Opus at a fraction of the cost.
StepFun releases Step-3.7-Flash, a new large vision-language MoE model with 198B parameters (11B active), 256K context, and up to 400 tokens/sec inference speed.