MiniMaxAI · video

MiniMax-H3

MiniMaxAI/MiniMax-H3

MiniMax-H3 at Q4_K_M is exactly 14,576,977,960 bytes (13.58 GiB / 14.58 GB) — an effective 3.521 bits per weight, not the nominal 4.

From the file· summed from 1 file(s)
Parameters
33.1B
Architecture
qwen3vl
Context
—
native (config.json)
License
other

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
Q4_010.60 GiB11,377,542,8802.748—molbal
Q2_K_M12.20 GiB13,102,161,0243.164—unsloth
Q2_K2 shards12.48 GiB13,402,362,0483.237—unsloth
Q5_012.93 GiB13,885,941,4723.354—molbal
Q4_K_M13.58 GiB14,576,977,9603.521—Abiray
Q3_K2 shards16.28 GiB17,475,434,6884.221—unsloth
Q4_K_M16.97 GiB18,218,065,0244.400—unsloth
Q8_019.94 GiB21,411,137,2485.171—molbal
Q4_K2 shards21.24 GiB22,801,760,4485.507—unsloth
Q5_02 shards25.90 GiB27,812,493,5046.717—unsloth
Q3_K_S2 shards29.00 GiB31,134,097,9847.520—Abiray
Q3_K_M2 shards29.00 GiB31,134,097,9847.520—Abiray
Q6_K2 shards30.87 GiB33,141,098,6888.004—unsloth
Q4_02 shards34.72 GiB37,279,210,0489.004—Abiray
Q4_K_S2 shards36.98 GiB39,706,667,5849.590—Abiray
Q4_K_M2 shards36.99 GiB39,717,542,0099.593—Abiray
Q8_02 shards39.91 GiB42,851,788,99210.350—unsloth
Q5_02 shards42.43 GiB45,558,594,11211.004—Abiray
Q5_K_S2 shards44.49 GiB47,774,968,38411.539—Abiray
Q5_K_M2 shards44.49 GiB47,774,968,38411.539—Abiray
Q6_K2 shards52.56 GiB56,438,100,48013.631—Abiray
Q8_02 shards67.12 GiB72,070,433,28017.407—Abiray

Pipeline components

a diffusion model is a graph of parts, not one file
ComponentSizeShareCan live on the CPU?
text encoder62.13 GiB47%yes — the usual first VRAM saving
denoiser61.73 GiB46%no, must be resident
vae9.70 GiB7%no, must be resident
Full pipeline133.56 GiBresident if nothing is offloaded

The parameter count published for a diffusion model describes the denoiser alone. Running it also requires its text encoder and VAE, and the text encoder is often nearly as large as the denoiser — which is why offloading it is the standard first move when you run out of memory.

We publish component sizes here, not throughput. Community-submitted image-generation rates do exist for many GPUs and we show them on the hardware pages, but they aggregate runs at different resolutions, step counts and settings, so they cannot be attributed to one model. Peak memory during sampling is unmeasured by any public source, and we do not estimate it.

Compare with

same modality, comparable size
vs LTX-2.3vs Wan2.2-I2V-A14Bvs LTX-2vs LTX2.3-10Erosvs Wan2.2-T2V-A14Bvs Wan2.2-Animate-14BAll alternatives →Will it fit on my card? →

Will it run on your card?

full quant x context sweep
By VRAM:8GB12GB16GB24GB32GB48GB
Radeon RX 6500 XT 4GBGeForce RTX 3050 6GBGeForce RTX 5050 8GBGeForce RTX 3080 10GBGeForce RTX 2080 Ti 11GBGeForce RTX 5070 12GBGeForce RTX 5060 Ti 16GBApple M3 Pro 18GBGeForce RTX 3080 Ti 20GBGeForce RTX 5090 D V2 24GBGeForce RTX 5090 D 32GBApple M5 Max 36GBApple M5 Pro 48GBApple M5 Max 64GBApple M3 Ultra 96GBApple M5 Max 128GBApple M2 Ultra 192GBApple M3 Ultra 256GBApple M3 Ultra 512GB

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 17.35 GiB. The real file is 13.58 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

Architecture unavailable — this repository is gated and no ungated mirror was found. Exact file sizes above are still authoritative; only the KV math needs the config.

Questions people ask

How much VRAM does MiniMax-H3 need?
Q4_K_M is exactly 14,576,977,960 bytes (13.58 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
Which quantization of MiniMax-H3 should I use?
Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.