One RTX 3090 runs a 27B agent plus image/video gen: real daily-use numbers
mightofkhan · reddit · 2026-11-17
The author runs a 27B LLM Discord agent and ComfyUI image/video generation on a single RTX 3090 (24GB), swapping models in and out of VRAM via a small custom MCP server. All numbers come from 205 real LLM requests and 35 real render jobs.
Key numbers
- Qwen3.8 27B (IQ4XS), 64k context: 72 tok/s avg (53–90), prompt processing 670–810 tok/s, MTP draft acceptance 61%, 22/24.6 GB VRAM
- SDXL images 32s end-to-end; 10s Wan 2.2 videos in 4–6 min, including 20s model swap
- Big agent job (40 model calls) 36 min; sustained load 85–90% GPU, 384W, max 67°C
Lessons learned
- IQ4XS + flash attention fits a 27B at 64k context in 24GB with 2.5GB to spare
- Use parallel 1, not 4 — same speed, extra slots risk VRAM spills
- 24GB fits the LLM or a render, not both; the 20s swap dominates quick images
- Long agent sessions slow down as the model re-reads full context each step — restart often
- Use Ethernet for always-on bots
Next experiment: q80 KV cache to free VRAM for longer context.
More from coding & agent
- A Pharma Agent That Flags Developability Risks on Every Sequence Registration — nlarusstone · 2026-11-17
- 55lb drone built with Claude Opus in 6 days — and then it crashed — IgorCarron · 2026-11-17
- Google ships Antigravity slash commands cheat sheet for modes, planning and automations — rseroter · 2026-11-17
- YC F26 Startup Maritime Launches a Managed Cloud for Running AI Agents — ycombinator · 2026-11-17
- Premiere AI Editing Test: Same 12-Min Interview Took 48 Min With AI vs 33 by Hand — BlackGull · 2026-11-17
- One-Sentence Prompt + Blender MCP Generates Isometric 3D Scenes Restylable at Will — rms80 · 2026-11-17