Moneta dual T4 — ComfyUI model study
Quality-first image generation on moneta (2× Tesla T4, 125 GB RAM) vs bigscreen
(RX 9070 XT, 31 GB RAM). Covers killing Trevor/Ollama, running ComfyUI on the T4s,
and which open models actually fit.
2×15 GBT4 VRAM (30 GB total)
125 GBMoneta system RAM
~16 GBBigscreen VRAM today
FP16Target precision on T4
Executive summary
ComfyUI on moneta's dual T4s works for quality — especially full
FP16 Flux and newer Klein 9B — but only after stopping Ollama (Trevor's 27B holds
~23 GB VRAM today). Dual T4 is not one 30 GB GPU by default; the win is
30 GB spread + 125 GB RAM for encoder offload. Bigscreen stays useful for fast
FP8 drafts; moneta becomes the quality backend.
Architecture today
Moneta · Hermies
→
Trevor → Ollama 27B (~23 GB VRAM)
Default / Isaac
→
Glifftop Ollama (unchanged)
Hermies comfy-gen
→
Bigscreen ComfyUI :8188
→
Flux FP8 + ~4.7 GB offload
| Machine | Role | Key resources |
|---|
| Moneta (10.0.0.10) |
Hermies hub, Immich, 70+ containers |
125 GB RAM, 2× T4 — Ollama uses ~23 GB VRAM |
| Glifftop (10.3.119.177) |
LLM for default Hermies + Isaac |
Separate Ollama — unaffected by moneta changes |
| Bigscreen (10.4.199.235) |
ComfyUI image gen today |
31 GB RAM, ~16 GB VRAM, Flux FP8 with partial offload |
Kill Trevor + Ollama?
What you lose
- Trevor persona (uncensored local 27B, Telegram on that profile)
- Direct
ollama CLI on moneta (~55 GB models on disk)
What you keep / gain
- Default Hermies + Isaac on glifftop — unchanged
- All
comfy-gen / Immich integration
- ~23 GB VRAM freed on both T4s + disk space back
Removing moneta Ollama does not improve bigscreen ComfyUI directly —
they're separate machines. It only enables ComfyUI on moneta.
Dual T4 — how ComfyUI uses VRAM
| Mode | Effective VRAM | Notes |
|---|
| Single GPU (default) |
~15 GB per job |
One T4; other idle or second instance on :8189 |
| CPU offload + one T4 |
15 GB GPU + 125 GB RAM |
Best path: UNet on GPU, T5-XXL on CPU — moneta's killer advantage |
| ComfyUI-MultiGPU / DisTorch2 |
~30 GB split |
cuda:0,50%;cuda:1,50% for large UNets |
T4 precision note
Tesla T4 (compute 7.5) has weak/no FP8 tensor cores. Prefer
FP16, BF16, or GGUF Q8 over FP8 for quality on T4.
FP8 is optimized for Ada/Hopper consumer GPUs.
Tier 1 — maximum quality (target these)
| Model |
VRAM |
Fits how |
Strength |
| Flux.1 Dev FP16 |
~24 GB UNet + ~10 GB T5 |
UNet split across T4s or T5 on CPU RAM |
Reference open-weight quality; huge LoRA ecosystem. Non-commercial. |
| FLUX.2 Klein 9B FP16 |
~27–29 GB |
Fills dual T4 at full precision |
Newest architecture; native img2img, multi-ref editing (8 refs). Non-commercial. |
| Krea 2 Raw BF16 |
~26 GB weights + encoder |
Dual T4 split + 125 GB RAM offload |
Best photorealism/aesthetic among 2026 open models. Community license. |
Tier 2 — excellent, single T4 (fully in VRAM)
Cleaner than bigscreen's current FP8 + 4.7 GB offload penalty.
| Model | VRAM | Notes |
|---|
| Flux.1 Dev Q8 GGUF |
~12–15 GB |
~95–98% of FP16; needs ComfyUI-GGUF node |
| Flux.1 Dev FP8 |
~12–14 GB |
Already on bigscreen; modest gain if no offload on moneta |
| Ideogram 4.0 FP8 |
~16–20 GB |
Best in-image text/typography; JSON prompts, not plain text |
| FLUX.2 Klein 9B FP8 |
~14–16 GB |
Editing + multi-ref; T5 offload brings to ~8–10 GB |
Tier 3 — strong but not peak quality
| Model | VRAM | Why not Tier 1 |
|---|
| Krea 2 Turbo FP8 |
~12 GB |
Distilled 8-step — fast and beautiful, lower ceiling than Raw |
| FLUX.2 dev GGUF Q4 |
~19 GB |
32B architecture but quantized; needs both T4s |
| Qwen-Image 2.0 FP8 |
~8–12 GB |
Good bilingual text + 2K; not best overall image quality |
| Z-Image Turbo |
~16 GB |
Speed-optimized, Apache 2.0 |
| SDXL + Juggernaut |
~6 GB |
On bigscreen today; fine for img2img, superseded for txt2img |
| FLUX.2 Klein 4B |
~8 GB |
Apache 2.0, editing — smaller model, lower ceiling |
Does not fit (even dual T4)
| Model | Why |
|---|
| FLUX.2 dev FP8 |
~32 GB — needs 32 GB+ single GPU |
| FLUX.2 dev BF16 |
~64 GB — datacenter only |
| HunyuanImage 3.0 |
40 GB+ |
| Wan 2.x 14B video (full quality) |
40 GB+ for 720p |
Recommendations by use case
| You want… | Best model on dual T4 |
|---|
| Overall max quality |
Flux.1 Dev FP16 (40–50 steps, 1024² or 1280²) |
| Newest architecture + editing |
FLUX.2 Klein 9B FP16 |
| Photorealism / aesthetics |
Krea 2 Raw BF16 |
| Text / typography in images |
Ideogram 4.0 FP8 |
| Simplest setup, still excellent |
Flux.1 Dev Q8 GGUF (one T4, LoRA headroom) |
| Commercial license |
FLUX.2 Klein 4B or Qwen-Image 2.0 (quality tradeoff) |
Recommended stack
- Flux.1 Dev FP16 — primary txt2img quality backend
- FLUX.2 Klein 9B FP16 — editing, multi-reference, img2img
- Ideogram 4.0 FP8 — anything with text in the image
- Flux.1 Dev Q8 GGUF — fallback on one T4 with LoRAs loaded
Skip for max quality: Krea 2 Turbo (distilled), Flux Schnell, Klein 4B, SDXL for new txt2img.
Proposed split
Moneta T4s (quality)
- Full Flux FP16, high steps, upscale chain
- ComfyUI on :8189 (or replace bigscreen as primary)
- CUDA PyTorch install (not ROCm)
- Optional ComfyUI-MultiGPU for dual-T4 workflows
Bigscreen (speed)
- FP8 quick drafts and iteration
- Or retire for images once moneta is stable
Disk budget (moneta has ~216 GB free)
| Model stack | Download size |
|---|
| Flux.1 Dev FP16 (full) | ~35 GB |
| FLUX.2 Klein 9B FP16 | ~25–30 GB |
| Krea 2 Raw BF16 | ~27 GB |
| Ideogram 4.0 FP8 | ~25–30 GB |
| Flux.1 Dev Q8 GGUF | ~13 GB |
Setup checklist
| Step | Action |
|---|
| 1 | Stop Ollama — frees both T4s |
| 2 | Install ComfyUI on moneta with CUDA PyTorch |
| 3 | Download model stack (~45–50 GB for Flux FP16 full) |
| 4 | Optional: ComfyUI-MultiGPU extension |
| 5 | Run on port 8189; point quality workflows at http://127.0.0.1:8189 |
| 6 | Disable Trevor profile in Hermies if Ollama is gone |
Research from Hermies infrastructure review · 2026-08-28 · Ignoring generation speed;
T4 throughput will be slower than RX 9070 XT but enables higher-precision models.
VRAM figures are practical estimates from ComfyUI community guides, not measured on this host yet.