Everything here is free and stays free — the format spec, the nodes, the workflows, the cartridges, the LoRAs. If it saved you a night of debugging (it contains several hundred of mine), tips keep the 5090 warm:<ul><li>☕ <a target="_blank" rel="ugc" href="https://ko-fi.com/joeygambino">Buy me a coffee on Ko-fi</a></li><li>💛 <a target="_blank" rel="ugc" href="https://github.com/sponsors/jlucasmcrell">Sponsor on GitHub</a></li><li>🔁 <a target="_blank" rel="ugc" href="https://liberapay.com/joeygambino">Liberapay</a> (recurring)</li></ul>⚡ Or right here: the Civitai tip button on this page sends Buzz directly. The <code>fl2va</code> half of the comfy-native set. These load with the built-in <code>UNETLoader</code> on stock ComfyUI 0.32+ — no custom node in the load path, no architecture patch, nothing to install first.Looking for <code>ref2va</code>? That is a separate listing: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2857809">MiniMax-H3 comfy-native</a>. Same formats, same method, same speeds — a different job, explained below.<h2 id="what-fl2va-does-that-ref2va-cannot">What fl2va does that ref2va cannot</h2><code>fl2va</code> is the first/last-frame variant. It carries no reference rows, so voice anchoring, the identity bank and reference images all do nothing on it. In exchange it does the one thing <code>ref2va</code> has no mechanism for: it lands on a supplied frame.<pre><code>checkpoint PSNR vs the supplied frame
--------------------------------- --------------------------
fl2va 26.35 dB
ref2va + keyframe, turbo 6 steps 16.15 dB
ref2va + keyframe, stock 20 steps 16.81 dB</code></pre>The stock-20-step row is the control: it rules out the sampler, so that ~10 dB gap is the checkpoint. <code>ref2va</code> treats a keyframe as a suggestion; <code>fl2va</code> treats it as the task. It will also take a first and a last frame and plan a camera move between them.Rule of thumb. ref2va when a face or a voice has to persist across shots. fl2va when a shot has to start exactly where the last one ended — stepping through a scene one shot at a time, feeding each shot's last frame into the next.This is not a size decision. fl2va and ref2va are byte-identical in size at every quantisation. "Lighter" only ever meant fewer tokens per sampling step, because there are no reference rows riding along.<h2 id="pick-by-your-gpu-not-by-the-file-size">Pick by your GPU, not by the file size</h2>A file that fits can still run emulated, and nothing in the UI tells you which you got. ComfyUI prints it at startup — find the line beginning <code>Native ops:</code> and check your format is on it.Emulation is hardware, not a ComfyUI version problem. NVFP4 and microscaling FP8 are Blackwell tensor-core features; <code>float8_e4m3fn</code> wants Ada or newer. An Ampere card emulates all three however current ComfyUI is.<pre><code>file GB take it if
------------------------------------------------ ------ ---------------------
MiniMax-H3-fl2va-pruned-comfy-nvfp4.safetensors 12.53 Blackwell, 50-series
MiniMax-H3-fl2va-pruned-comfy-w4a8.safetensors 12.54 Ada, 40-series
MiniMax-H3-fl2va-pruned-comfy-int8.safetensors 20.97 Ampere, 30-series
MiniMax-H3-fl2va-pruned-comfy-fp8.safetensors 20.96 Ada+, 24 GB and up
MiniMax-H3-fl2va-pruned-comfy-mxfp8.safetensors 21.56 Blackwell, 24 GB up</code></pre>On a 4090 a user measured <code>w4a8</code> at 17.4 s/it against 31.0 s/it for <code>fp8</code> on identical settings — because fp8 at 21 GB does not fit beside the activation pool on 24 GB and streams weights over PCIe. Native ops you cannot reach are worth nothing.<h2 id="all-of-them-beat-the-gguf">All of them beat the GGUF</h2>Measured on the ref2va set — same architecture, same layer set, same quantiser, so it carries over. Same scene, same seed, same shape: 544×960, 243 frames, 2 shots, 10 steps, so the weights are the only variable. RTX 5090, against the Q8_0 GGUF.<pre><code>build GB shot 1 shot 2 vs Q8_0
------------- ------- ---------- ---------- ----------------
nvfp4 12.53 12.71 s/it 19.57 s/it -34% / -28%
int8 20.97 13.5 20.1 -30% / -26%
mxfp8 21.56 13.4 20.9 -31% / -23%
w4a8 12.54 14.05 20.50 -27% / -25%
fp8 20.96 14.8 22.4 -24% / -18%
Q8_0.gguf 21.50 19.36 27.18 (baseline)</code></pre><code>int8</code> is the row worth staring at: effectively the same file size as the Q8_0 GGUF and 30% quicker. That gap is dequantisation overhead in the GGUF path, not a difference in quantisation quality.<h2 id="they-are-all-the-same-quality-as-each-other">They are all the same quality as each other</h2>Worth saying plainly, because it saves you a download. The same scene was rendered through <code>w4a8</code> (12.54 GB) and <code>int8</code> (20.97 GB) at 704×1280 and 20 steps, and both were reviewed as video by the same reviewer with the same prompt. They produce the same defects: garbled text on instrument dials, reflective safety-tape edges that warp during movement, limb distortion on fast motion, stiff lip sync.So paying 8 GB more does not buy better output. Those defects are characteristics of the base model at these shapes — a prompting and settings problem, not a weights problem, and no quantisation here moves them.Take the smallest file your card runs natively. The larger ones are not a quality upgrade. And if you hit those artifacts, do not report them as quantisation damage — check against the unquantised model or a GGUF first, because you will very likely find them there too.<h2 id="two-formats-i-built-and-then-threw-away">Two formats I built and then threw away</h2><ul><li><code>convrot_w4a4</code> — renders without error and the output is visibly degraded. It quantises activations to 4 bits; every file here keeps activations at 8. Instrument text becomes scribble, hi-vis banding dissolves, hands smear. Caught by render at seed 4040 against four other builds. It saves 1.2 GB against <code>w4a8</code> and costs all of it.</li><li><code>float8_e5m2</code> — loads cleanly, then throws at the first matmul: Multiplication of two Float8_e5m2 matrices is not supported. Same on LTX-2.5. It would ship as a file that looks fine until you sample.</li></ul>Neither was built for fl2va. Both failures are properties of the format, not of these particular weights, so building 32 GB in order to quarantine it would have been busywork.<h2 id="why-nobody-has-k-quants-of-this-model">Why nobody has K-quants of this model</h2>H3's hidden width is 2688, and K-quants need weight rows divisible by 256. 2688 is not, so the whole K-ladder is unavailable for the DiT — which is why the GGUFs of it are Q4_0 / Q5_1 / Q8_0 and never Q4_K_M. The encoder is a stock Qwen3-VL and its rows do divide, so that repo has real K-quants; the two are easy to confuse.The comfy-native formats have no such constraint, which is part of why this set exists.<h2 id="how-they-were-made">How they were made</h2>The layer set is mirrored from Comfy-Org's own <code>minimax_h3_fl2va_pruned_int8_convrot</code> rather than guessed: exactly 200 Linears — <code>attn.qkv_proj</code>, <code>attn.out_proj</code>, <code>mlp.fc1</code> and <code>mlp.fc2</code> across 50 blocks. Everything else stays bf16/fp16: adaLN, the timestep embedders, every norm and bias, the token refiner, the patch projections, the final layer. Those steer every block, and rounding them is how a quantised DiT dies.Built from the pruned bf16 lineage, which matters arithmetically. On the unpruned lineage adaLN is 38% of the file and 4-bit lands at 19.9 GB — what the existing GGUF Q4_0 already achieves. Pruned 4-bit lands near 12.5 GB.How these were verified, honestly. These were forward-gated, not render-tested: every build ran a GPU forward pass on an attention layer and a feed-forward layer, and all five passed. That catches a format that cannot execute; it does not catch one that executes and looks wrong. The speed table, the quality equivalence and both cuts are inherited from the ref2va set, which was rendered at matched seed and shape on identical architecture. If you render a fl2va build and it disagrees with any of this, please say so.<h2 id="what-else-you-need">What else you need</h2>The text encoder and the two VAEs from <code>Comfy-Org/MiniMax-H3</code> (video and audio are separate). Nothing else — that is the point of these.<h2 id="honest-limits">Honest limits</h2><ul><li>No render A/B on the fl2va weights specifically — the findings above are inherited from ref2va, not measured here.</li><li><code>fl2va</code> ignores <code>voice_ref</code>, <code>self_anchor_voice</code> and the identity bank entirely. If your workflow leans on those, you want ref2va.</li><li>Resolution cannot change mid-chain, and neither variant changes that.</li></ul><h2 id="support">Support</h2>Everything I publish is free and stays free. If it saved you a night of debugging, tips keep the 5090 warm: <a target="_blank" rel="ugc" href="https://ko-fi.com/joeygambino">Ko-fi</a> · <a target="_blank" rel="ugc" href="https://github.com/sponsors/jlucasmcrell">GitHub Sponsors</a> · <a target="_blank" rel="ugc" href="https://liberapay.com/joeygambino">Liberapay</a>.

Files with SHA256: 18ce05a9e588618db5ae6dca380f242bf8041d9d47f3ffb5cd80d9dec9f9b1ec