{
  "id": 15907444,
  "title": "Four big local LLMs on one RTX 5090 and 128 GB of RAM, benchmarked side by side",
  "url": "https://urgent.news/2026/10/21/four-big-local-llms-on-one-rtx-5090-and-128-gb-of-ram-benchmarked",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-21T15:20:49.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/seppegadeyne/four-big-local-llms-on-one-rtx-5090-and-128-gb-of-ram-benchmarked-side-by-side-2h8j"
  },
  "original_language": "en",
  "account": "One high-end workstation, equipped with an RTX 5090 GPU and 128 GB of RAM, was used to benchmark four large language models (LLMs) running simultaneously. The models, ranging from 125B to 748B parameters, were tested on three different runtimes over a single day on October 21, 2026. The article presents the side-by-side results, showcasing how each setup performed.\n\nThe four models were Qwen3.8-Flash-Next with 125B parameters, Ornith-1.5-122B with 122B parameters, GLM-5.3-Flash with 321B parameters, and DeepSeek-V4.1-Flash-1017 with 748B parameters. All four models are mixture-of-experts models, meaning only a fraction of the model is activated for each token, allowing them to run on a single GPU.\n\nThe RTX 5090 GPU features 32 GB of VRAM and drives the desktop display. The CPU used was a Ryzen 9 9950X3D with 16 cores and 32 threads. The system was running Arch Linux (Omarchy), NVIDIA driver 615.78, and CUDA 13.4. The operating environment kept 3 GB of RAM free for each runtime, ensuring no interference from other processes.\n\nTo measure performance, each model was tested individually with no other model runtime or GPU miner running. Benchmarking consisted of decoding speed, time to first token, prefill speed, and drafting quality. Decode speed was measured using a client-side streaming benchmark against the OpenAI endpoint, with eight runs alternating between a Dutch prose prompt and a coding prompt. Prefill speed was calculated as the number of prompt tokens divided by the total request time. Drafting speed was evaluated using speculative decoding by default for three of the four models.",
  "summary": "One RTX 5090, 128 GB of RAM, and four open models from 125B to 748B parameters. Over the past weeks I got each of them running on the same workstation, on three different runtimes. On October 21, 2026 I measured all four on the same day, one after the other, with the same benchmarks. This article puts the numbers side by side, so you can see what each setup is good for. Every number below was…",
  "key_points": [
    "All four models are mixture-of-experts, allowing them to run on a single GPU with limited VRAM.",
    "Performance measured by decoding speed, time to first token, prefill speed, and drafting quality."
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}