Model Library

The MiMo V2.6 family from Xiaomi introduces omnimodal Mixture-of-Experts models in Flash and Pro variants at 309B and 1.02T parameters. Trained through unified reinforcement learning, both support 1M-token context for coding, computer-use, and cybersecurity agents.

The VultronCoderAtlas family from Vultr introduces a 2.1T-parameter sparse Mixture-of-Experts model derived from Kimi K3. It targets coding, multimodal understanding, tool use, and long-context agentic workflows with efficient expert activation.

The Atria Dawn family from InternLM introduces an agentic model built on the 744B-parameter GLM-5.2 foundation. It targets scientific automation, software creation, research, document generation, cybersecurity, and long-horizon workflows with iterative tool use, 1M-token context, and sparse MoE architecture.

The Intern S2 family from InternLM introduces a 397B multimodal foundation model designed for scientific intelligence, general reasoning, and long-horizon agents. It combines sparse MoE architecture, hybrid linear and full attention, 262K-token context, and vision capabilities for advanced scientific and agentic workflows.

The Ling 3.0 family from InclusionAI introduces efficient hybrid reasoning model designed for production-scale agentic workflows. It emphasizes inference efficiency, strong reasoning and instruction following, long-context capabilities, and reliable execution across coding, general, and deep-research tasks.

The DeepSeek V4.1 family from DeepSeek AI introduces multimodal Mixture-of-Experts model built for long-context reasoning and agentic workloads. It combines million-token contexts, efficient sparse architectures, controllable reasoning, and native vision-language capabilities.

The Nex-N2.5 family from Nex AGI introduces next-generation agentic models spanning mini, Pro, and Max variants. It advances computer use, web browsing, visual grounding, and long-horizon task execution, with the Max model built on a 1.6T-parameter MoE foundation.

The IFM K2 Horizon family introduces six open models spanning 0.9B to 375B parameters, unified across reasoning, coding, mathematics, and agentic tasks. Its fully open training lifecycle enables reproducibility, research, and flexible deployment from edge devices to enterprise systems.

The DeepSeek V4 family from DeepSeek AI introduces large-scale Mixture-of-Experts models with up to trillion-parameter capacity and 1M-token context support. It incorporates hybrid attention, advanced optimization, and domain-specialized training to enable efficient long-context reasoning and strong multi-domain performance.

The Hy4 Preview family from Tencent introduces a frontier-class Mixture-of-Experts model with 770B parameters and 49B active parameters. It emphasizes advanced reasoning, million-token context, sparse attention, and efficient agentic workloads.

The GLM-5.3 family from Zhipu AI introduces natively multimodal model combining sparse and linear attention with MoE architecture. It targets efficient long-context reasoning, coding, and advanced agentic workflows with million-token context.

The Qwen3.8 family from Alibaba Cloud introduces a frontier-class open model designed for coding, professional work, research, and long-horizon agentic tasks. It combines flexible reasoning control with extended context support up to 1M tokens.

The Granite 4.2 family from IBM introduces dense reasoning models spanning 3B to 30B parameters, designed for coding, mathematics, tool calling, agentic workflows, multilingual dialogue, and efficient long-context enterprise applications.

The Ornith 1.5 family from DeepReinforce.AI introduces self-improving foundation models through an end-to-end reinforcement learning loop. It spans 397B and 35B MoE and 9B dense models, combining self-generated tasks, task-specific scaffolds, and solution rollouts to strengthen reasoning, coding, and agentic capabilities across diverse workflows.

The LFM2.5 VL family from Liquid AI introduces compact multimodal models optimized for efficient on-device vision-language inference. It spans multiple sizes supporting image understanding, OCR, grounding, multilingual applications, and function calling.

The Nemotron 3 family is a series of efficient open language models developed by NVIDIA, built on a hybrid Mixture-of-Experts architecture. The models emphasize high token throughput, long-context reasoning, and cost-efficient deployment for agentic and enterprise AI workloads.

The Muse Glimmer family from Meta introduces compact multimodal language model optimized for local agentic deployment. It combines reasoning, tool use, failure recovery, and multimodal understanding to support reliable autonomous workflows on consumer hardware.

The LFM2.5 family from Liquid AI introduces compact hybrid models optimized for efficient on-device language, reasoning, tool use, and agentic workloads. It spans 230M to 8.3B parameters, combining convolution and GQA attention architectures with extended context support up to 131K tokens.

The Inkling family from Thinking Machines introduces open-weight multimodal language model designed for reasoning, coding, and agentic applications. It combines native text, image, audio, and video understanding with efficient Mixture-of-Experts architecture to support conversational, tool-use, and long-context workflows.

The Kimi K3 family from Moonshot AI introduces next-generation multimodal Mixture-of-Experts model built for frontier-scale reasoning, coding, and agentic workflows. As the world's first open 3T-class model family, it combines native multimodal understanding, million-token context, and advanced attention architectures for efficient long-horizon performance.