← Topper Bowers

Self-hosted MiniMax H3 vs. fal.ai H3 Max

Same prompts, same seed, three renders per test: self-hosted H3 in draft (cheap iteration tier) and full (delivery tier) on a rented Nebius H200 interruptible VM, against fal's H3 Max. The VM costs $2.45/hr, and a cold boot takes about 45s. Run 2026-08-27. Generation time and cost below are the render itself from a warm worker, excluding the VM's cold-start time.

Self-hosted setup

The self-hosted side runs on one Nebius H200 interruptible VM: one GPU, 16 vCPUs, 200 GiB of memory, and the NVIDIA H200 NVLink platform. The queue is nebius-h200.

It runs MiniMax H3, with fl2va loaded by default. That is the first-and-last-frame-to-video-and-audio variant, using the pruned, int8, convrot-quantized checkpoint minimax_h3_fl2va_pruned_int8_convrot.safetensors. ref2va is the single-reference-image variant; both loads both checkpoints.

Attention uses NVIDIA's Sol-Attn kernel, not Sage. It is wired as a self-attention-only override through ComfyUI's optimized_attention_override, targeting H3's head_dim=128. The first 20% of denoising steps and first two transformer blocks stay on dense attention. Dispatch counting makes a silent dense fallback visible. When Sol-Attn is off, the runtime falls back to attention_backend() (usually SDPA or Flash, depending on the environment).

Modes

Mode controls the speed/quality trade-off. The variant is separate.

still
One frame at the requested canvas; scene plates, up to about 1MP.
draft
Half canvas, turbo transformer, 8 steps; iteration and animatics.
draft_refine
Full canvas with an LTX-2.5 pass; previews and sharing, not photoreal delivery.
turbo
Full canvas, turbo transformer, 8 steps; useful for stylized or animated work.
full
Full canvas, base transformer, 20 steps; the delivery tier.

Variants

Variant controls the input condition, independently of the mode.

fl2va
First-and-last-frame conditioning with native video and audio output.
ref2va
Single reference-image conditioning with native video and audio output.
both
Loads both checkpoints so either conditioning path is available.
Summary — both tests, three renders each
Metric Self-hosted (draft) Self-hosted (full) fal H3 Max
Test 1 — T2V, 10s cost $0.0297 $0.286 $0.40 promo / $0.80 full
Test 2 — I2V, 5s cost $0.0160 $0.047 $0.20 promo / $0.40 full
Test 1 — generation time 41.4s 398.6s 13s wall / 7.8s inference
Test 2 — generation time 22.3s 65.1s 5s wall / 1.5s inference
Output resolution 672×384 (T2V) / 448×576 (I2V) 1344×768 (T2V) / 768×864 (I2V) 1344×768 (T2V) / 768×864 (I2V)

Test 1 — Text-to-video, native audio, 10s

seed 42 · both platforms, identical prompt
Prompt: Realistic cinematic aerial establishing shot at golden hour: a small songbird flies gracefully over the Eiffel Tower in Paris, sunlight glinting off the iron lattice, shallow depth of field, photoreal film grain. [Push in] The bird descends and swoops close to a young woman standing on a nearby rooftop terrace; she smiles, looks up at the bird, and says warmly, “Hi bird!”
Self-hosted H3 · Nebius H200 · draft mode Self-hosted H3 · Nebius H200 · full mode fal.ai H3 Max · 768P tier
Self-hosted (draft)
672×384
Generation time
41.4s
Cost
$0.0297
Self-hosted (full)
1344×768
Generation time
398.6s
Cost
$0.286
fal H3 Max (768P)
1344×768
Generation time
13s
Cost, promo
$0.40 through Aug 31
Cost, full rate
$0.80 from Sep 1

Test 2 — Image-to-video, first frame, 5s

seed 42 · first frame from projects/jesus-knocks reference/characters/boy
Reference: boy-greeting-wave-up.jpg as first frame.
Prompt: [Push in] The boy keeps waving, then a small bird flies down and lands close beside him; he grins, looks up at the bird, and says warmly, “Hi bird!”
Self-hosted H3 · Nebius H200 · draft mode Self-hosted H3 · Nebius H200 · full mode fal.ai H3 Max · 768P tier
Self-hosted (draft)
448×576
Generation time
22.3s
Cost
$0.0160
Canvas
896×1152 → 448×576 closest valid portrait on draft's grid
Self-hosted (full)
768×864
Generation time
65.1s
Cost
$0.047
Canvas
768×864 full mode isn't grid-locked, matches source exactly
fal H3 Max (768P)
768×864
Generation time
5s
Cost, promo
$0.20 through Aug 31
Cost, full rate
$0.40 from Sep 1
Canvas
768×864 matches source image exactly
Draft mode's canvas grid is rigid — its no-duplication floor and its base-canvas cap coincide exactly, leaving only nine valid canvases total, all sharing one pixel budget. 896×1152 → 448×576 (ratio 0.78) is the closest of those nine to the source image's 0.89, not an exact match. fal isn't grid-constrained the same way, so it reproduces the source ratio exactly.
draft is the cheap iteration tier on the self-hosted side by design (half-canvas, 8 steps) — fal has no equivalent cut-rate mode to compare against. full (20 steps, full canvas) is self-hosted's actual delivery tier, and the gap to fal narrows a lot there: 8.6× cheaper than fal's full rate at this resolution instead of draft's 12.5×, at roughly 13× fal's generation time instead of draft's 4.5×.