Local LLMs

Strata Models and Memory Requirements: Which to Pick on a 12GB or 16GB VRAM GPU

Choose Qwen3.8-Flash-Next, Gyro, Coder or DeepSeek V4.1-Flash on Strata by system RAM and 12GB/16GB VRAM. Includes DeepSeek's memory and disk needs and second-GPU results.
2026.11.06 AI Hardware Zukan
Local LLMs
Local AI

What Changed in Strata Updates: Should 12GB and 16GB GeForce Users Upgrade?

Strata changes up to 0.2.10 for 12GB/16GB GeForce on Windows: upgrading, changed defaults and how to revert, and 0.1.41 vs 0.1.45 speed on RTX 5080/5060 Ti.
2026.11.06 AI Hardware Zukan
Local AILocal LLMs
Local AI

Using a Second GPU in Strata: Layer Split vs. Helper Cache on RTX 5080 + 5060 Ti

Layer split vs. helper cache in Strata 0.1.39–0.1.41 on an RTX 5080 with an OCuLink RTX 5060 Ti. Every setup tried had slower prompt processing than one card.
2026.11.06 AI Hardware Zukan
Local AILocal LLMs
ComfyUI

ComfyUI Upscale Eats 67GB RAM? Fix with External ESRGAN

ComfyUI's 4K upscale consumes 67-77GB of system RAM. Learn how to run ESRGAN via spandrel for constant ~1.4GB usage and 2x speedup, complete with code.
2026.06.15 AI Hardware Zukan
ComfyUI
GPUs & Graphics Cards

Why Oculink GPU Fails in Ollama: Dual Setup & Scheduling

Fix RTX 5060 Ti recognition in Ollama via Oculink. Step-by-step guide using CUDA_VISIBLE_DEVICES and OLLAMA_SCHED_SPREAD for RTX 5080 + RTX 5060 Ti dual-GPU setups.
2026.06.15 AI Hardware Zukan
GPUs & Graphics CardsLocal AIMini PCs & Small Form Factor PCs
GPUs & Graphics Cards

RTX 4070 Super vs RTX 5060 Ti: VRAM Showdown for LLMs

Upgrading from RTX 4070 Super to RTX 5060 Ti boosts 14B model speed. See how the 4GB VRAM difference impacts performance across 15 models in our real-world test.
2026.06.15 AI Hardware Zukan
GPUs & Graphics CardsMini PCs & Small Form Factor PCs
ComfyUI

ComfyUI Multi-GPU Guide: RTX 5080 + 4070 Super Real-World Tests

Real-world ComfyUI dual-GPU guide (RTX 5080 + 4070 Super): port separation for parallel jobs, why VRAM can't be pooled, and where Ollama is the exception.
2026.06.14 AI Hardware Zukan
ComfyUIGPUs & Graphics CardsMini PCs & Small Form Factor PCs
GPUs & Graphics Cards

RTX 5080 MoE Models: Power Drops to One-Quarter in Real-World Tests

RTX 5080 across 14 LLMs: dense models draw 200–300W, but MoE (gemma4:26b, qwen3.5:35b-a3b) drop to 47–73W at 14.8GB VRAM. With tokens/W and power-cost analysis.
2026.06.14 AI Hardware Zukan
GPUs & Graphics Cards
Local AI

Qwen3.6-35B-A3B: Local MoE Multimodal LLM Guide

Run Qwen3.6-35B-A3B locally on RTX 5090. Explore specs, FP8 model size, and multimodal capabilities of this efficient MoE LLM.
2026.06.13 AI Hardware Zukan
Local AI
GPUs & Graphics Cards

Anima TrainFlow: Train LoRA on 6GB VRAM

Anima TrainFlow is a Web trainer for Anima 2B LoRA on 6GB+ VRAM GPUs. Features sd-scripts + Gradio + Prodigy, live preview, portable distribution, and VRAM guidelines.
2026.06.13 AI Hardware Zukan
GPUs & Graphics Cards
Next page