MiniMax · Over 80B · Mixture of experts
MiniMax M3
427.04 billion parameters in total, but only 25.86 billion are read for each token. That split is the whole point of the design: it costs the memory of a large model and the speed of a small one. It uses grouped-query attention with 4 KV heads for 64 query heads, which cuts the cache to 6% of what full attention would need.
About MiniMax M3
MiniMax M3 is a mixture-of-experts model with 427.04B parameters, 25.86B of them active for each token, built from 60 layers with a hidden size of 6144. Its repository was published on 2026-06-02. The license listed on Hugging Face is other. It uses grouped-query attention with 4 KV heads for 64 query heads, which cuts the cache to 6% of what full attention would need.
Memory needed, by quantization
At an 8k context window. Quality is our rough ranking of how much the compression costs you: anything at or above Q5 is hard to tell apart from the original in normal use.
| Quantization | File size | KV cache | Total VRAM | Quality | What it costs you |
|---|---|---|---|---|---|
| Q8_0 | 422.57 GB | 0.94 GB | 422.76 GB | 99% | Lossless in practice. Use it when the memory is there. |
| Q6_K | 326.12 GB | 0.94 GB | 326.6 GB | 98% | Very close to Q8 for two thirds of the size. |
| Q5_K_M | 282.87 GB | 0.94 GB | 283.47 GB | 96% | The quality-first choice when Q6 will not fit. |
| Q4_K_M | 240.12 GB | 0.94 GB | 240.84 GB | 93% | The default. Best size-to-quality ratio for local use. |
| IQ4_XS | 211.28 GB | 0.94 GB | 212.09 GB | 91% | Importance-matrix 4-bit. Q4_K_S quality, smaller file. |
| Q3_K_M | 194.38 GB | 0.94 GB | 195.23 GB | 86% | Degradation starts to show. A way to fit one size up. |
| IQ3_XXS | 152.12 GB | 0.94 GB | 153.1 GB | 79% | Aggressive. Only worth it on very large models. |
What a longer conversation costs
Same model at Q4_K_M, only the context window changes. The cache grows in a straight line with every token in the window.
| Context | KV cache | Total VRAM | Fits in 8 GB | Fits in 12 GB | Fits in 24 GB |
|---|---|---|---|---|---|
| 4k | 0.47 GB | 240.37 GB | no | no | no |
| 8k | 0.94 GB | 240.84 GB | no | no | no |
| 16k | 1.88 GB | 241.78 GB | no | no | no |
| 32k | 3.75 GB | 243.65 GB | no | no | no |
| 64k | 7.5 GB | 247.4 GB | no | no | no |
| 128k | 15 GB | 254.9 GB | no | no | no |
The "fits" columns allow for the roughly 0.8 GB Windows keeps for the desktop.
Which hardware runs MiniMax M3
0 of 118 consumer devices run it at a quantization worth using, at an 8k context window.
| Device | Memory | Verdict | Quantization | Used | Speed |
|---|---|---|---|---|---|
| Apple M1 Ultra 128GB Apple | 128 GB | Just fits | n/a ! | 103.79 GB | 44.3 t/s |
| Apple M2 Ultra 128GB Apple | 128 GB | Just fits | n/a ! | 103.79 GB | 44.3 t/s |
| Apple M2 Ultra 192GB Apple | 192 GB | Just fits | n/a ! | 153.51 GB | 31.2 t/s |
| Apple M4 Max 128GB Apple | 128 GB | Just fits | n/a ! | 103.79 GB | 30.2 t/s |
| Apple M3 Ultra 256GB Apple | 256 GB | Just fits | n/a | 212.67 GB | 23.6 t/s |
| Apple M3 Max 128GB Apple | 128 GB | Just fits | n/a ! | 103.79 GB | 22.1 t/s |
| NVIDIA DGX Spark 128GB NVIDIA | 128 GB | Just fits | n/a ! | 103.82 GB | 16.4 t/s |
| Apple M3 Ultra 512GB Apple | 512 GB | Just fits | n/a | 423.95 GB | 12.2 t/s |
| Ryzen AI Max+ 395 128GB AMD | 128 GB | Just fits | n/a ! | 103.82 GB | 11.8 t/s |
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Partial offload 60/60 | n/a | 240.84 GB | 6 t/s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Partial offload 60/60 | n/a | 240.84 GB | 4.5 t/s |
| GeForce RTX 5090 NVIDIA | 32 GB | Partial offload 60/60 | n/a | 240.84 GB | 4.4 t/s |
| RTX A6000 NVIDIA | 48 GB | Partial offload 60/60 | n/a | 240.84 GB | 4.4 t/s |
| Radeon PRO W7900 AMD | 48 GB | Partial offload 60/60 | n/a | 240.84 GB | 4.3 t/s |
| GeForce RTX 4090 NVIDIA | 24 GB | Partial offload 60/60 | n/a | 240.84 GB | 4.1 t/s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Partial offload 60/60 | n/a | 240.84 GB | 4.1 t/s |
| GeForce RTX 3090 NVIDIA | 24 GB | Partial offload 60/60 | n/a | 240.84 GB | 4 t/s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Partial offload 60/60 | n/a | 240.84 GB | 4 t/s |
| RTX A5000 NVIDIA | 24 GB | Partial offload 60/60 | n/a | 240.84 GB | 4 t/s |
| Radeon RX 7900 XTX AMD | 24 GB | Partial offload 60/60 | n/a | 240.84 GB | 4 t/s |
| GeForce RTX 5080 NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.9 t/s |
| GeForce RTX 5070 Ti NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.9 t/s |
| GeForce RTX 3080 Ti NVIDIA | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.9 t/s |
| GeForce RTX 3080 12GB NVIDIA | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.9 t/s |
| GeForce RTX 5080 Laptop NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.9 t/s |
Show 35 more devices
| Radeon RX 7900 XT AMD | 20 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.9 t/s |
| GeForce RTX 5070 NVIDIA | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.8 t/s |
| GeForce RTX 4080 SUPER NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.8 t/s |
| GeForce RTX 4080 NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.8 t/s |
| GeForce RTX 4070 Ti SUPER NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.8 t/s |
| GeForce RTX 3080 10GB NVIDIA | 10 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.8 t/s |
| GeForce RTX 2080 Ti NVIDIA | 11 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.7 t/s |
| GeForce RTX 4090 Laptop NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.7 t/s |
| Radeon RX 9070 XT AMD | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.7 t/s |
| Radeon RX 9070 AMD | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.7 t/s |
| Radeon RX 7900 GRE AMD | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.7 t/s |
| Radeon RX 7800 XT AMD | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.7 t/s |
| GeForce RTX 5060 Ti 16GB NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| GeForce RTX 4070 Ti NVIDIA | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| GeForce RTX 4070 SUPER NVIDIA | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| GeForce RTX 4070 NVIDIA | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| GeForce GTX 1080 Ti NVIDIA | 11 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| GeForce RTX 4080 Laptop NVIDIA | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| RTX A4000 NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| Radeon RX 6900 XT AMD | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| Radeon RX 6800 AMD | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| Arc A770 16GB Intel | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.6 t/s |
| GeForce RTX 3060 12GB NVIDIA | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.5 t/s |
| Radeon RX 7700 XT AMD | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.5 t/s |
| GeForce RTX 2060 12GB NVIDIA | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.4 t/s |
| Radeon RX 6700 XT AMD | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.4 t/s |
| Arc B580 Intel | 12 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.4 t/s |
| GeForce RTX 4060 Ti 16GB NVIDIA | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.3 t/s |
| Arc B570 Intel | 10 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.3 t/s |
| Radeon RX 7600 XT AMD | 16 GB | Partial offload 60/60 | n/a | 240.84 GB | 3.2 t/s |
| GeForce RTX 5060 Ti 8GB NVIDIA | 8 GB | Partial offload 1/60 | n/a | 240.84 GB | 2 t/s |
| GeForce RTX 5060 NVIDIA | 8 GB | Partial offload 1/60 | n/a | 240.84 GB | 2 t/s |
| GeForce RTX 4060 Ti 8GB NVIDIA | 8 GB | Partial offload 1/60 | n/a | 240.84 GB | 2 t/s |
| GeForce RTX 4060 NVIDIA | 8 GB | Partial offload 1/60 | n/a | 240.84 GB | 2 t/s |
| GeForce RTX 3070 Ti NVIDIA | 8 GB | Partial offload 1/60 | n/a | 240.84 GB | 2 t/s |
How it compares
Other models in the same family or of a similar size, on a 12 GB card and a 24 GB card, each at the best quantization that fits.
| Model | Weights at 4 bits | GeForce RTX 3060 12GB | GeForce RTX 4090 |
|---|---|---|---|
| MiniMax M3 | 240.1 GB | does not fit well | does not fit well |
| MiniMax M2.5 | 128.6 GB | does not fit well | 10.4 t/s at Q3_K_M |
| MiniMax M2.7 | 128.6 GB | does not fit well | 10.4 t/s at Q3_K_M |
| GLM 4.5 | 201.5 GB | does not fit well | does not fit well |
What to buy to run MiniMax M3
The cheapest hardware that runs it at a quantization worth using and a speed you would not resent, at an 8k context window.
Cheapest that works
Apple M3 Ultra 256GB
$5,599 whole machine
IQ4_XS · uses 212.67 GB of its 256 GB · about 23.6 tokens/s
Best value
Apple M3 Ultra 512GB
$9,499 whole machine
Q6_K · about 15.7 tokens/s · 819 GB/s
If it just has to run
GeForce RTX 2060 12GB
$160 used
Q3_K_M · about 4.2 tokens/s. Cheaper because it runs the model at a lower precision.
| Device | Price | Memory | Runs it at | Speed | |
|---|---|---|---|---|---|
| Apple M3 Ultra 256GB Apple · whole machine | $5,599 | 256 GB | IQ4_XS | 23.6 t/s | Check price |
| Apple M3 Ultra 512GB Apple · whole machine | $9,499 | 512 GB | Q6_K | 15.7 t/s | Check price |
Indicative prices reviewed 2026-09-01; used prices are marketplace typical. Options that cost more than a cheaper one with no more memory and no more speed are hidden. Price links are Amazon affiliate links. Change the standard, the context window or the budget in the buying tool.
Where it sits in the catalog
Source: MiniMaxAI/MiniMax-M3 on Hugging Face. Downloaded 196k times in the last month. Published 2026-06-02. Architecture figures are read from the repository's own configuration file, so they move when the model does.
Questions people ask about MiniMax M3
How much VRAM does MiniMax M3 need?
At Q4_K_M with an 8k context window, MiniMax M3 needs about 240.84 GB: 239.43 GB of weights, 0.94 GB of KV cache and 0.47 GB of overhead. At the highest-quality file it needs about 422.76 GB.
What is the cheapest GPU that runs MiniMax M3?
No consumer device in our catalog fits this model entirely.
How much does a longer context window add for MiniMax M3?
On the memory side, a 32k window adds 3.8 GB of KV cache on top of the weights, and the cache grows in a straight line with every token. The longest window the model supports natively is 1M tokens.