Tagged "moe-architecture"
6 articles tagged moe-architecture, 15 April 2026 to 4 October 2026. Newest first.
-
Aleph Alpha Releases Kolibri: A 78.1B Open-Weight English-German MoE Model
Aleph Alpha has released Kolibri, a 78.1B Mixture-of-Experts model with only 3.46B active parameters, enabling efficient local deployment of high-capacity multilingual models with minimal compute requirements.
-
Show HN: I shrank DeepSeek V4 Flash to 57GB and it wrote a compiler on my Mac
A developer successfully compressed DeepSeek V4 Flash to 57GB and demonstrated its capability to write a compiler on a Mac. This showcases practical quantization and model optimization techniques for running state-of-the-art models on consumer hardware.
-
Ollama Releases NVIDIA Nemotron 3.5 Lightning for Agent Execution
Ollama v0.32.9 now includes NVIDIA's Nemotron 3.5 Lightning, a 30B MoE model with only 3B active parameters optimized for on-device agent execution. This lightweight model is designed for frameworks like OpenClaw and Hermes Agent, making powerful agentic AI accessible on local hardware.
-
Thinking Machines Lab Releases Inkling-Small: A 276B Total, 12B Active Open Weights Multimodal MoE Model
Thinking Machines Lab has released Inkling-Small, an open-weights multimodal mixture-of-experts model with 276B total parameters but only 12B active during inference, enabling efficient local deployment on consumer hardware.
-
Tencent Open-Sources Hy3 295B MoE Model Built for STEM Reasoning
Tencent releases Hy3, a 295B mixture-of-experts model optimized for STEM reasoning tasks. This open-source release provides local LLM practitioners with a high-capacity model option for specialized reasoning workloads.
-
Dynamic Expert Cache in llama.cpp Achieves 27% Faster Inference on Large MoE Models
A new optimization technique for llama.cpp improves CPU+GPU token generation speed by 27% on Qwen3.5-122B through dynamic expert caching, raising practical inference rates from 15 to 23 tokens per second.