Can MiniMax M2.7 run on NVIDIA H20 96GB?

YES — With Q2_K

A80Great

Estimated from fit model

MiniMax M2.7 needs ~104.0 GB VRAM. NVIDIA H20 96GB has 96.0 GB. With Q2_K quantization, expect ~75 tok/s.

Runtime: llama.cppCapacity: OffloadBandwidth: HighStack: StandardBottleneck: Host offload

Operating mode

Choose the run profile you care about

Interactive favors responsiveness, while light API and scale-out lean harder on serving readiness. The fit stays the same, but the recommendation lens changes.

Current mode

Balanced

Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.

MiniMax M2.7 at UD-IQ4_XS needs 154.6 GB — too much for NVIDIA H20 96GB (96.0 GB). Runs at Q2_K (104.0 GB) with low quality.

Capabilities:

Select quantization to explore

F16 (Maximum quality) — 485.8 GB, exceeds 96.0 GB available

485.8 GB required96.0 GB available

506% VRAM needed

389.8 GB over capacity — needs offload or smaller quantization

Fit status

Too heavy

Decode

4.7 tok/s

TTFT

41301 ms

Safe context

Memory

485.8 GB / 96.0 GB

Offload

80%

Memory breakdown

Weights471.5 GB

KV Cache3.8 GB

Runtime0.9 GB

Headroom9.6 GB

See how fast it feels

With memory offload — actual speed may be lower

See how fast it feelsMiniMax M2.7 on NVIDIA H20 96GB
1st promptCold start — includes initialization
>Write a Python function to sort a list
Here's a Python function that sorts a list using quicksort:

def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[len(arr) // 2]
    left   = [x for x in arr if x < pivot]
    middle = [x for x in arr if x == pivot]
    right  = [x for x in arr if x > pivot]
    return quicksort(left) + middle + quicksort(right)

# Example usage
numbers = [3, 6, 8, 10, 1, 2, 1]
print(quicksort(numbers))  # [1, 1, 2, 3, 6, 8, 10]

The algorithm picks a pivot, partitions the list into smaller, equal, and larger elements, then recursively sorts each partition and concatenates the results.
2nd promptWarm — model ready, no init overhead
>Now add type hints and a docstring
def quicksort(arr: list[int]) -> list[int]:
    """Sort a list using the quicksort algorithm.

    Args:
        arr: List of integers to sort.

    Returns:
        A new sorted list.
    """
    if len(arr) <= 1:
        return arr
    pivot = arr[len(arr) // 2]
    left   = [x for x in arr if x < pivot]
    middle = [x for x in arr if x == pivot]
    right  = [x for x in arr if x > pivot]
    return quicksort(left) + middle + quicksort(right)

Added type hints for the input and return type, plus a Google-style docstring.
Estimated: 4.7 tok/s decode · 41.3s TTFT (warm) · 12 tok/s prefill

What limits this setup

It fits through host-memory offload, and offload is the main reason performance drops.

CPU or host-memory offload is active

About 10% of the working set spills out of accelerator memory, which usually hurts latency and sustained decode throughput.

Very little memory headroom

You can run the model, but there is not much room left for longer context, bigger batches, extra apps, or future model updates.

Best improvement path

Remove offload with more accelerator memory

Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.

Buy headroom, not only minimum fit

A slightly larger memory tier gives you safer context growth and makes the recommendation more future-proof.

Increase host RAM if you keep offloading

This setup may need roughly 6.9 GB of extra host RAM just for the offloaded portion, before OS and other tools.

Performance by workload

Workload	Grade	Fit	Decode	TTFT	Context
Chat	F	Too heavy	29.8 tok/s	3549 ms	4K
Coding	F	Too heavy	29.2 tok/s	6641 ms	4K
Agentic Coding	F	Too heavy	28.0 tok/s	10052 ms	4K
Reasoning	F	Too heavy	29.2 tok/s	7848 ms	4K
RAG	F	Too heavy	28.0 tok/s	12565 ms	4K

Inference speed

MiniMax M2.7 inference speed — tokens per second by GPU & Mac

Estimated decode speed (tokens/sec) for MiniMax M2.7 at UD-IQ4_XS across popular GPUs and Apple Silicon, including multi-GPU rigs, using the fastest local runtime per device. Fastest is Mac Studio M3 Ultra 256GB at ~20 tok/s. Speed is memory-bandwidth bound, so cards that fit the whole model in VRAM run far faster than ones that offload to system RAM.

GPU / Mac	Memory	Quant	Speed (tok/s)	Fits?
Mac Studio M3 Ultra 256GB	256 GB	UD-IQ4_XS	20.0	Tight
Mac Studio M2 Ultra 128GB	128 GB	UD-IQ4_XS	8.3	Too big
Mac Studio M1 Ultra 128GB	128 GB	UD-IQ4_XS	7.9	Too big
MacBook Pro M4 Max 128GB	128 GB	UD-IQ4_XS	6.2	Too big
MacBook Pro M4 Max 64GB	64 GB	UD-IQ4_XS	5.6	Too big
RTX 5090 32GB	32 GB	UD-IQ4_XS	4.2	Too big
2× RX 7900 XTX 24GB	48 GB	UD-IQ4_XS	4.1	Too big
MacBook Pro M3 Max 64GB	64 GB	UD-IQ4_XS	3.9	Too big
MacBook Pro M1 Max 64GB	64 GB	UD-IQ4_XS	3.6	Too big
MacBook Pro M4 Pro 48GB	48 GB	UD-IQ4_XS	3.4	Too big
2× RTX 4090 24GB	48 GB	UD-IQ4_XS	2.8	Too big
RTX 4090 24GB	24 GB	UD-IQ4_XS	2.7	Too big
RX 7900 XTX 24GB	24 GB	UD-IQ4_XS	2.4	Too big
2× RTX 3090 24GB	48 GB	UD-IQ4_XS	2.4	Too big
RTX 3090 24GB	24 GB	UD-IQ4_XS	2.3	Too big
RTX 4080 Super 16GB	16 GB	UD-IQ4_XS	2.1	Too big
4× RTX 3060 12GB	48 GB	UD-IQ4_XS	2.1	Too big
RTX 4070 12GB	12 GB	UD-IQ4_XS	2.0	Too big
RTX 3060 12GB	12 GB	UD-IQ4_XS	2.0	Too big
RTX 4060 8GB	8 GB	UD-IQ4_XS	2.0	Too big

Estimates for single-stream decoding at UD-IQ4_XS; real tokens/sec varies with prompt length, context, batch size, and runtime build. Prompt processing (prefill) is faster than the decode figures shown here.

Quantization options

How MiniMax M2.7 (230B params) fits at each quantization level on NVIDIA H20 96GB (96.0 GB usable).

Quant	Bits	VRAM	Quality	Fit
Q2_K	2	89.7 GB	Low	F0
Q3_K_S	3	112.7 GB	Low	F0
NVFP4	4	128.8 GB	Medium	F0
Q4_K_M	4	140.3 GB	Medium	F0
Q5_K_M	5	165.6 GB	High	F0
Q6_K	6	188.6 GB	High	F0
Q8_0	8	246.1 GB	Very High	F0
F16	16	471.5 GB	Maximum	F0

Get started

Copy-paste commands to run MiniMax M2.7 on your machine.

Run

lms load MiniMax-M2.7 && lms server start

Opções de upgrade

Hardware que roda bem MiniMax M2.7

NVIDIA B200 180GBOpção econômica

180 GB VRAM (+84)8000 GB/s (+4000)

Makes the model fit on the accelerator instead of staying completely out of reach.155.6 tok/s decodificação

Makes the model fit on the accelerator instead of staying completely out of reach.

Removes host-memory offload, which is usually the single biggest latency and throughput win.

~$30,000 MSRP

NVIDIA H200 141GBMelhor custo-benefício

141 GB VRAM (+45)4800 GB/s (+800)

Makes the model fit on the accelerator instead of staying completely out of reach.65.2 tok/s decodificação

Makes the model fit on the accelerator instead of staying completely out of reach.

Raises estimated decode speed by about 123%.

~$30,000 MSRP

NVIDIA H200 PCIe 141GBUpgrade NVIDIA

141 GB VRAM (+45)4800 GB/s (+800)

Makes the model fit on the accelerator instead of staying completely out of reach.55.8 tok/s decodificação

Makes the model fit on the accelerator instead of staying completely out of reach.

Raises estimated decode speed by about 91%.

~$30,000 MSRP

Frequently asked questions

Can NVIDIA H20 96GB run MiniMax M2.7?

Yes, NVIDIA H20 96GB can run MiniMax M2.7 at Q2_K quantization (Very compromised (needs ~6.9 GB host RAM)). The recommended UD-IQ4_XS requires 154.6 GB which exceeds available memory, but at Q2_K it needs only 104.0 GB. Expected decode speed: 74.6 tok/s.

How much VRAM does MiniMax M2.7 need?

MiniMax M2.7 (230B parameters) requires approximately 154.6 GB at UD-IQ4_XS quantization. On NVIDIA H20 96GB, it fits at Q2_K using 104.0 GB.

What is the best quantization for MiniMax M2.7?

The recommended quantization is UD-IQ4_XS, but on NVIDIA H20 96GB the best fitting quantization is Q2_K, which uses 104.0 GB.

What speed will MiniMax M2.7 run at on NVIDIA H20 96GB?

On NVIDIA H20 96GB, MiniMax M2.7 achieves approximately 74.6 tokens per second decode speed with a time-to-first-token of 2597ms using Q2_K quantization.

Can NVIDIA H20 96GB run MiniMax M2.7 for coding?

For coding workloads, MiniMax M2.7 on NVIDIA H20 96GB receives a F grade with 29.2 tok/s and 4K context.

What context window can MiniMax M2.7 use on NVIDIA H20 96GB?

On NVIDIA H20 96GB, MiniMax M2.7 can safely use up to 4K tokens of context at Q2_K quantization. The model's official context limit is 205K, but available memory constrains the safe maximum.

What should I upgrade first if MiniMax M2.7 feels slow on NVIDIA H20 96GB?

Remove offload with more accelerator memory. Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.

See all results for NVIDIA H20 96GB See all hardware for MiniMax M2.7

Embed this result▼

Paste this snippet into any page to show a live fit card.

<iframe src="https://willitrunai.com/embed/minimax-m2-7-on-h20-96gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>

Preview: