Open models, served at the edge

Harmonizes high computational efficiency with superior reasoning and agent performance.

1.6T-parameter flagship MoE (49B active) for maximum reasoning, coding, and agentic performance.

Cost-efficient V4 variant — 284B parameters (13B active) with the same 1M-token context.

A coding partner built for core coding, vibe coding, tool use, and complex reasoning.

744B-parameter model targeting complex systems engineering and long-horizon agentic tasks.

MIT-licensed 744B MoE with a 1M-token context — tops open-weight coding and design benchmarks.

OpenAI's open-weight model for powerful reasoning, agentic tasks, and versatile developer use.

Native multimodal agentic model with instant and thinking modes.

Roughly 1T-parameter MoE with 256K context, tuned for long-horizon tasks and multi-step tool use.

Coding-specialized K2 variant that cuts thinking tokens by about 30% versus K2.6.

Natively multimodal mixture-of-experts model — 17B parameters across 128 experts.

Instruction-tuned 70B model optimized for multilingual dialogue.

RL-trained across real-world environments — SOTA in coding, agentic tool use, and search.

First open-weight model combining frontier coding, 1M context, and native multimodality.

Open hybrid Mamba-Transformer MoE — 120B parameters (12B active) with a native 1M-token context.

Unified vision-language foundation with an efficient hybrid architecture.

A smaller but faster version of the Qwen3.5 397B A17B model.

Coding-focused Qwen3 variant with advanced agentic capabilities and IDE integration.