MiniMax H3 local render times swing fourfold on identical GPUs, dev.to roundup finds

A single GPU, a fourfold spread

A post on dev.to consolidates MiniMax H3 timings and VRAM reports from about 20 community threads, all from the past month, and its headline finding is counterintuitive: the spread in render times within a single GPU model is larger than the spread between different cards. One RTX 5090 owner reported roughly 22 minutes for 362 frames at 768×1024 using 8 steps and the Turbo sampler, while other 5090 users replied that they finish similar work in 3 to 5 minutes. According to the post, both claims are true, and the difference comes down to optimisation rather than hardware.

The collected figures span a wide range of setups. On the 5090 alone, the roundup lists 10 minutes 16 seconds at 896×1120 without turbo at 20 steps, an unspecified "under 5 minutes", 134 seconds for a 15-second clip, and about 11 minutes for a 90-second video using a 12 GB-optimised workflow. Further down the stack, an RTX 3090 reportedly manages around 3 minutes with a 6-step turbo LoRA and attention optimisation; a 5070 Ti paired with 64 GB of system RAM recorded 118 seconds cold and 93 seconds warm at 608×352 (230 seconds at 480×864); and a 4070 with the same RAM finished the default 608×352 workflow in 167 seconds.

Four settings explain most of the variance

After reading the reply threads, the author identifies four recurring factors behind the spread. First, Turbo LoRA with step counts cut from 20 to 4–6, described as the single largest factor. Second, the attention backend, with SageAttention repeatedly named as the fix for slow times. Third, SLA speedup — one reply attributes the 22-minute figure directly to not using it. Fourth, resolution and total megapixels: 608×352 versus 896×1120 represents a large difference in work, and people often quote times without stating either.

The post's advice to anyone benchmarking is to always report resolution, frame count, step count and Turbo LoRA status, because half the numbers circulating are unusable without them.

The VRAM floor is lower than assumed, with a caveat

According to the roundup, 12 GB of VRAM is enough to run H3. There are working reports on a 3080 12 GB with 32 GB of system RAM, and a purpose-built low-VRAM workflow produced 30 seconds of output in roughly 14 minutes on a 12 GB card. The caveat is that system RAM matters as much as VRAM here: the working 12 GB reports are paired with 32–64 GB of RAM, and a 5060 Ti with 16 GB of VRAM and 128 GB of RAM still hit out-of-memory on the default text-to-video template. The post's conclusion is that the default template is not tuned for mid-range cards, and switching to a low-VRAM workflow is what actually lowers the floor.

The 30-second generation is stitched

Worth knowing before planning around it: the 30-second low-VRAM workflow relies on comfyui-h3-multishot, which joins three 10-second clips. The post says the seams are handled, but it is not a native single-pass generation, and anyone needing one continuous take should test it before committing.

The economics of "$0 API cost"

The roundup also pushes back on "$0 API cost" framing, noting a top reply to the original 5090 post observing that the hardware costs plenty before you ever press the queue button. The author offers a break-even formula — hardware cost divided by seconds per clip times the API rate per second — and, while disclosing an affiliation with the company, quotes hosted H3 rates of $0.074 per second at 768P and $0.119 per second at 2K. At those rates a 10-second 768P clip costs about $0.74, meaning a $2,000 card pays for itself somewhere near 2,700 clips, before electricity and waiting time.

On where each approach wins, the post is blunt: local generation suits high volume, constant iteration, custom LoRAs, absence of content restrictions and tolerance for 3-to-20-minute waits per attempt. Hosted APIs suit low or bursty volume, immediate needs, cards below the memory floor, and 2K output where local memory pressure becomes severe.

Why it matters

H3 is new enough that scattered anecdotes are the main practical guidance available, and this consolidation makes three things concrete. Configuration, not hardware, dominates local video generation performance — tuning step counts and the attention backend can matter more than a GPU upgrade. The realistic entry point is 12 GB of VRAM plus 32–64 GB of system RAM with a tuned workflow, not the default template. And comparable benchmarks require full settings disclosure; the post's closing call for timings with resolution, frames, steps and attention settings highlights the discipline the community currently lacks. For borderline buyers wondering whether a mid-range card qualifies, the break-even arithmetic often suggests an API is cheaper than new hardware.