Tag: #moe-models14 results
Open-weight 27B-32B models tested on Terminal-Bench 2.0, with Qwen 3.6-27B scoring 38.2% under default timeouts. This performance aligns with late 2025 hosted frontier models like Claude Opus 4.1. MOE models show significantly faster inference speeds.
智源 Flag OS 社群宣布完成 Alibaba Qwen3.8-2.4T-A95B 在九款 AI 晶片上的 Day0 適配、精度對齊與部署驗證。新版本支援 BF16、FP8 與 INT8，讓開發者能在多種国产與主流 AI 硬體上快速部署這款 2.4T 參數 MoE 模型。
Comprehensive benchmarks of 37 LLMs on MacBook Air M5 32GB using llama-bench with Q4_K_M quantization. Highlights MoE models like Qwen 3.5 35B-A3B at 31 tok/s as top performers. Open-source tool shared for community hardware benchmarks.
Gemma 4 is the first LLM to hit 100% success in user's multilingual tool calling tests across English, German, and Japanese. Used in N8N voice assistant on 68GB VRAM setup with MoE models like Gemma4 26BA4B.
Cursor's Composer 2 uses Moonshot's open-source Kimi K2.5 model via Fireworks AI, sparking discussions on China's AI supply chain. DeepSeek and Kimi models are increasingly foundational for global apps like OpenClaw agents. This shift mirrors manufacturing supply chains, with tokens as the new infrastructure.
Detailed breakdown of China's LLM scene featuring ByteDance's Doubao as proprietary leader and Alibaba's Qwen excelling in open-weight models. Deepseek innovates with MLA and other tech, while 'Six Small Tigers' like Zhipu and Minimax release large open-weight MoE models for cheap inference. Meituan pushes aggressive open releases like 562B LongCat-Flash.
DeepSeek announces V4 multimodal model using fully domestic chips, bypassing Nvidia. Chinese firms optimize algorithms like MoE to cut costs and advance国产 chips for training. This marks a shift from inference to full training on non-Nvidia hardware.
AWS implemented multi-LoRA inference for Mixture of Experts (MoE) models in vLLM, including kernel-level optimizations. This enables efficient serving of dozens of fine-tuned models on Amazon SageMaker AI and Amazon Bedrock. GPT-OSS 20B serves as the primary example.
Benchmark pits Google's new Gemma4 against Alibaba's Qwen3.5 for local agentic coding on RTX 4090. Qwen3.5-27B excels in code quality, reliability, and efficiency despite slower speeds. MoE variants like Gemma4-26B are faster but less accurate on complex tasks.
6-hour benchmarks on V100 32GB GPU across 20 LLM models (MoE & dense) with Ryzen 7600X, testing power limits from 300W-150W and CPU offload levels up to 32K context. Power limiting to 200W yields <2% loss for generation; MoE models handle offload far better than dense. Top performers: Nemotron-30B Mamba2 at 152 t/s.