Kimi K2.5
A large language model in the Kimi series developed by Moonshot AI. It supports image understanding capabilities, enabling users to generate code directly by uploading screenshots, design mockups, and similar visual materials.
Core Facts
Timeline (last 90 days)
AI coding software Goeser confirmed in a technical report that Goeser 2 continues training based on Kimi K2.5.
The judging panel runs on Kimi K2.5, completing a review in about every 9 seconds, with a 90% agreement rate with human reviewers' judgments.
Running Kimi K2.5 on an NVIDIA B200 with real coding agent trajectories (SWE-smith, EvalScope evaluation), TokenSpeed outperforms TensorRT-LLM across the entire latency/throughput curve; at batch size 1 it is about 9% faster, and at about 100 tokens/second per user its throughput is about 11% higher.
When serving Kimi K2.5 at 100 tokens/second, Jalapeño's throughput is close to 9 times that of the second-place competitor, with a peak throughput of up to 700 tokens/second.
The Jalapeño benchmark simultaneously ran three models employing different MoE and attention mechanisms: GPT-OSS 120B, DeepSeek R1, and Kimi K2.5.
The user is considering writing their own agent orchestrator, switching to Kimi K2.5, or building a model-agnostic architecture.
All Facts (9)
K2.5的Agent集群采用主控Agent（Orchestrator）负责任务拆解与结果聚合、子Agent（Worker）执行原子化任务的架构
75%UnverifiedK2.5的MoE架构单次推理实际激活的参数量可能只有总参数的1/8甚至更少
70%Partially VerifiedKimi K2.5是大规模混合专家（MoE）架构的语言模型
65%UnverifiedAI coding software Goeser confirmed in a technical report that Goeser 2 continues training based on Kimi K2.5.
50%UnverifiedThe judging panel runs on Kimi K2.5, completing a review in about every 9 seconds, with a 90% agreement rate with human reviewers' judgments.
50%UnverifiedRunning Kimi K2.5 on an NVIDIA B200 with real coding agent trajectories (SWE-smith, EvalScope evaluation), TokenSpeed outperforms TensorRT-LLM across the entire latency/throughput curve; at batch size 1 it is about 9% faster, and at about 100 tokens/second per user its throughput is about 11% higher.
50%UnverifiedWhen serving Kimi K2.5 at 100 tokens/second, Jalapeño's throughput is close to 9 times that of the second-place competitor, with a peak throughput of up to 700 tokens/second.
50%UnverifiedThe Jalapeño benchmark simultaneously ran three models employing different MoE and attention mechanisms: GPT-OSS 120B, DeepSeek R1, and Kimi K2.5.
50%UnverifiedThe user is considering writing their own agent orchestrator, switching to Kimi K2.5, or building a model-agnostic architecture.
50%