DeepSeek V4 Pro 0813 is an MoE (Mixture-of-Experts) model with a 1M-token context window, released Aug 12, 2026. This guide explains MoE, tokens, long-context tradeoffs, and how to call the model on OpenRouter with reasoning-enabled streaming.
DeepSeek V4 Flash 0731 is a MoE model release evaluated on ARC-AGI-2 with multiple reasoning-effort modes. It reports 61.4% on ARC-AGI-2 Semi-Private at $0.04/task (max effort), alongside efficiency-focused design like long-context attention and DSpark speculative decoding.
Swiftlet demonstrates how to run Qwen MoE models locally by keeping only the dense core in RAM and streaming routed experts from SSD on demand. It reports ~4.3GB RAM for a 4-bit 80B model on an M5 Mac, and an on-device 35B experience on iPhone around ~2.5GB RAM.
TurboFieldfare runs Gemma 4 26B-A4B on Apple Silicon with ~2GB RAM by keeping core weights + an FP16 KV cache resident and streaming only the routed MoE experts from SSD. This post explains the memory pieces (MoE routing, KV cache, quantization) and walks through building, installing, and tuning the runtime.
Kimi K3 is Moonshot AI’s open 2.8T-parameter multimodal model with a 1M-token context window. Built on KDA and Attention Residuals, plus stable MoE expert routing, it’s positioned as a frontier-capable open-weight system—especially for long-horizon coding and even GPU kernel/compiler workflows.
Thinking Machines’ Inkling is an open-weights, multimodal MoE LLM with 975B total parameters and 41B active per token. It supports long context (up to 1M tokens), controllable thinking effort, and fine-tuning via Tinker.