
Built on the architectural foundation of Qwen3.5, running on EAGLE V4 speculative decoding.

NVIDIA’s largest and final Nemotron 3 model, a hybrid Mamba-Attention MoE with 55B active of 550B total parameters, built for frontier-scale reasoning, tool use, and agentic tasks.

The retrained July 31 checkpoint of DeepSeek V4-Flash (284B total / 13B active), with gains coming from improved post-training for coding, agents, reasoning, and tool use.

Tencent’s 770B MoE (49B active, 1M context) built for long-horizon agentic work like coding, game dev, and complex document analysis.

DeepSeek’s cost-efficient V4.1 tier, a sparse mixture-of-experts model that DeepSeek says outperforms its previous V4 Pro flagship, with native text and image support and a 1M-token context window.

Z.ai’s flagship coding and agentic model, post-trained on the GLM-5.2 base with a 1M-token context, text-only, and reasoning always on.

Z.ai’s first natively multimodal GLM-5 model, a 320B-total / 18B-active MoE with a 1M-token context window, tuned for efficient coding and long-horizon agent tasks.

StepFun’s flagship agentic model, a sparse MoE with 27B active out of 600B total parameters, aimed at software engineering and professional knowledge work, especially finance.

MiniMax’s multimodal coding and agentic model built on its Sparse Attention architecture, supporting up to 1M tokens of context plus image and video input.

Moonshot AI’s most capable model, a 2.8T-parameter open-weight multimodal MoE with a 1M-token context, designed for long-horizon coding, knowledge work, and reasoning.

Xiaomi’s low-cost omni-modal reasoning model with 1M context, built for high-volume workloads and accepting text, images, video, and audio as input.