MiniMax H3 is an open-weight video-plus-audio generator, not a chat LLM. Its core is a 33.1B-parameter dense transformer (66.3 GB of BF16 weights). Text goes through a full Qwen3-VL-32B encoder (66.7 GB in BF16). With both VAEs, keeping every component resident is about 144 GB of weights. It is not MoE, MiniMax publishes no token context length, and there is no LLM-style KV cache to size. This page gives weight arithmetic only. Activations, attention buffers and VAE decode add more, and those were not measured here.
Key specs from official sources
| Spec | Value | Source |
|---|---|---|
| Model type | Omni-modal generator: text, image, video and audio in; video with stereo audio out | Model card |
| Architecture | H3-Omni-Transformer, dense single-stream (not MoE) | Model card, “H3-Omni-Transformer” |
| Transformer params | 33B stated; safetensors metadata total 33,122,992,896 | Model card; HF API |
| Active params | Same as total (dense). About 13B are AdaLN branches that need not be loaded for inference only | Model card |
| Text encoder | Full Qwen3-VL-32B, hidden states from layer 50 | Model card, “H3-Encoder” |
| Context length | Not published in tokens. Output 4–15 s, 24 FPS, 32 kHz stereo audio | Model card |
| Local resolution | H3-Base generates 768p. 2K needs the hosted H3-Regenerate-2K | Model card |
| Weights | Open, BF16, MiniMax H3 Community License | Model card; open-source notice |
| Release | Announced 2026-07-31; weights open-sourced 2026-08-03 | MiniMax blog; open-source notice |
VRAM table: weights by precision
GB is 10⁹ bytes. BF16 rows use the real safetensors file sizes at revision 42ed227. FP8, Q8 and Q4 rows are 1 byte and 0.5 byte per parameter. No official FP8, Q8 or Q4 checkpoint exists. Those rows are arithmetic, and real 4-bit formats add a few percent for scales.
| Component | Params | FP16/BF16 | FP8 | Q8 (8-bit) | Q4 (4-bit) |
|---|---|---|---|---|---|
| Omni transformer, full | 33.1B | 66.3 GB | ≈33.1 GB | ≈33.1 GB | ≈16.6 GB |
| Omni transformer, AdaLN dropped (≈20B, per model card) | ≈20B | ≈40 GB | ≈20 GB | ≈20 GB | ≈10 GB |
| Qwen3-VL-32B text encoder | ≈33.4B (from file size) | 66.7 GB | ≈33.4 GB | ≈33.4 GB | ≈16.7 GB |
| Visual VAE + audio VAE (as shipped) | not stated | 11.0 GB | — | — | — |
| Everything resident, full transformer | ≈144 GB | ≈78 GB | ≈78 GB | ≈44 GB |
The FP8 and Q4 totals keep the VAEs at their shipped 11.0 GB. The ≈20B AdaLN-dropped row uses MiniMax’s “approximately 13B” figure, so it is rounded. The encoder runs once per request, so pipelines can load it, encode the prompt, and free it before denoising. That is why one transformer plus VAE can fit where the full 144 GB cannot.
Token counts instead of KV cache
H3 denoises the whole clip at once. Memory grows with the number of video tokens, not with a cached chat history. From the model card, the visual VAE compresses 16× in space and 4× in time, and the 1×2×2 patch makes the transformer see 32× spatial and 4× temporal downsampling. Audio is 40 latent tokens per second per channel. At 1344×768 and 24 FPS that gives roughly:
| Clip length | Frames | Video tokens (≈42 × 24 × frames/4) | Audio tokens per channel |
|---|---|---|---|
| 5 s | 120 | ≈30,000 | 200 |
| 10 s | 240 | ≈60,500 | 400 |
| 15 s (max) | 360 | ≈90,700 | 600 |
These are rounded counts from published compression factors. Exact padding and frame rules were not checked. The transformer hidden size was not taken from a config here, so this page does not convert tokens into GB. The released code uses full attention only. MiniMax says sparse attention will ship later, which should cut long-clip cost.
GPU recommendations
| Setup | What is known | Fit |
|---|---|---|
| 1× 24 GB (RTX 4090 / 4090 D) | Upstream SGLang report: about 18 GB peak, 303 s for 107 frames at 1344×768, INT8 DiT and layerwise offload | Works only with offload. Needs plenty of host RAM and patience |
| 1× 32 GB (RTX 5090) | No report here. Q4 or FP8 DiT alone (10–33 GB) still needs encoder offload | Untested |
| 1× 80 GB (H100 / A100 80GB) | BF16 transformer (66.3 GB) fits alone. Encoder must be offloaded or run first | Tight. Weight math only |
| 128 GB unified (DGX Spark GB10) | Upstream vLLM-Omni report: 97.7 GiB allocator peak with FP8 DiT, about 36 minutes per request | Works, slow |
| 4× 80 GB or larger | The model card’s SGLang example uses --num-gpus 4 with Ulysses parallelism |
Fastest documented path |
Multiple discrete GPUs do not pool into one big VRAM number. The four-GPU setup splits sequence work across cards. See two GPUs and unified memory and file size is not runtime VRAM.
How to run MiniMax H3 locally
Download from Hugging Face with hf download MiniMaxAI/MiniMax-H3. Scope it to FL2VA/* for text-to-video and first/last-frame work, or Ref2VA/* for reference-driven work. MiniMax lists SGLang, vLLM-Omni, diffusers and ComfyUI as supported runtimes. The hosted H3-Context-IR prompt rewriter is not open-sourced. MiniMax strongly recommends it or its prompting guide for output quality. Without the hosted H3-Regenerate-2K, local output stops at 768p.
The weights are under the MiniMax H3 Community License. The license excludes some territories and puts a revenue threshold on commercial use. Read it before you deploy.
What this page does not claim
The calculator on this site models LLMs, so H3 has no calculator preset and no will-it-fit page. Nothing here was measured by this site. Activation memory, VAE decode peaks and the visual VAE’s parameter count and dtype were not checked. The 4090 D and GB10 numbers are upstream reports.
FAQ
How much VRAM does MiniMax H3 need?
Holding every BF16 component resident at once is about 144 GB of weights: 66.3 GB transformer, 66.7 GB text encoder, 10.4 GB visual VAE and 0.6 GB audio VAE, before activations. With layerwise offload, an upstream SGLang report ran 768p text-to-video-plus-audio on one 24 GB RTX 4090 D at about 18 GB peak, much slower.
Is MiniMax H3 a mixture-of-experts model?
No. The official model card describes the H3-Omni-Transformer as a 33B-parameter dense, single-stream Transformer. Total and active parameters are the same. About 13B of those parameters sit in AdaLN branches that MiniMax says need not be loaded for inference-only deployment.
What is the MiniMax H3 context length?
MiniMax does not publish a context length in tokens. H3 is a diffusion video model, not an autoregressive LLM, so it has no KV cache that grows with a chat context. The published limits are output length of 4 to 15 seconds, 24 FPS, a default 768-pixel short side for H3-Base, and up to 12 reference files for Ref2VA.
Are the MiniMax H3 weights open?
Yes. MiniMax open-sourced the weights on 2026-08-03 on Hugging Face as MiniMaxAI/MiniMax-H3, under the MiniMax H3 Community License, not Apache or MIT. The license has territory exclusions and a revenue threshold; read it before deploying. The hosted H3-Context-IR and H3-Regenerate-2K modules are not in the release.
Can I run MiniMax H3 on a 24 GB GPU?
Only with offload. The upstream SGLang cookbook reports one 24 GB RTX 4090 D finishing a 1344x768, 107-frame T2VA run in about 303 seconds with INT8 DiT linears and layerwise streaming of the DiT and text encoder. The weights do not fit resident in 24 GB at any listed precision.
Are there official FP8 or GGUF Q4 builds of MiniMax H3?
Not in the official repository. The released checkpoints are BF16. The FP8, Q8 and Q4 rows on this page are bytes-per-parameter arithmetic, not downloadable files. Runtimes such as SGLang and vLLM-Omni apply their own on-the-fly quantization to the DiT.