Search Results for "moe"

MiniMax-M2 is an open-weight large language model designed specifically for high-end coding and agentic workflows while staying compact and efficient. It uses a Mixture-of-Experts (MoE) architecture with 230 billion total parameters but only 10 billion activated per token, giving it the behavior of a very large model at a fraction of the runtime cost. The model is tuned for end-to-end developer flows such as multi-file edits, compile–run–fix loops, and test-validated repairs across real repositories and diverse programming languages. ...

Qwen3-Coder is the latest and most powerful agentic code model developed by the Qwen team at Alibaba Cloud. Its flagship version, Qwen3-Coder-480B-A35B-Instruct, features a massive 480 billion-parameter Mixture-of-Experts architecture with 35 billion active parameters, delivering top-tier performance on coding and agentic tasks. This model sets new state-of-the-art benchmarks among open models for agentic coding, browser-use, and tool-use, matching performance comparable to leading models...

Ornith-1.0 is a large open-source reasoning model from DeepReinforce, built for agentic coding, tool use, and complex software engineering workflows. It is part of the Ornith 1.0 family, which includes dense and MoE models post-trained on Gemma 4 and Qwen 3.5. The model focuses on coding-agent performance across benchmarks such as Terminal-Bench, SWE-Bench, NL2Repo, OpenClaw, and ClawEval. Its training uses a self-improving reinforcement learning framework that optimizes not only solution attempts but also the scaffolds that guide those attempts, helping the model discover better search paths and produce higher-quality solutions. ...

LongCat-2.0 is Meituan’s flagship open-weight Mixture-of-Experts language model designed for frontier-scale coding, reasoning, and autonomous agent workflows. It features 1.6 trillion total parameters with approximately 48 billion activated per token, combining high capability with efficient sparse inference. The model was pretrained on more than 35 trillion tokens and trained entirely on a large-scale cluster of domestically developed AI accelerators, demonstrating stable frontier-scale...

Kimi K2.7 Code is a coding-focused agentic model built on Kimi K2.6, designed for long-horizon software engineering, autonomous coding workflows, and complex tool-based execution. It improves end-to-end task completion across real-world programming scenarios while reducing thinking-token usage by about 30% compared with K2.6. Architecturally, it uses a 1T-parameter Mixture-of-Experts design with 32B activated parameters, 61 layers, 384 experts, a 256K-token context window, and a MoonViT...