Can DeepSeek V4 Flash run on AMD Instinct MI300X 192GB?

YES — Tight Fit

S96Excellent

Estimated from fit model

DeepSeek V4 Flash needs ~179.4 GB VRAM. AMD Instinct MI300X 192GB has 192.0 GB. With NVFP4 quantization, expect ~89 tok/s.

Runtime: llama.cppCapacity: TightBandwidth: HighStack: StandardBottleneck: Balanced

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.

Capabilities:

Select quantization to explore

NVFP4 (Medium quality) — 179.4 GB, 89.1 tok/s, Tight fit

179.4 GB required192.0 GB available

93% VRAM used

Fit status

Tight fit

Decode

89.1 tok/s

TTFT

2174 ms

Safe context

169K

Memory

179.4 GB / 192.0 GB

Memory breakdown

Weights158.0 GB

KV Cache1.3 GB

Runtime0.9 GB

Headroom19.2 GB

See how fast it feels

See how fast it feelsDeepSeek V4 Flash on AMD Instinct MI300X 192GB
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: 89.1 tok/s decode · 2.2s TTFT (warm) · 223 tok/s prefill

What limits this setup

This setup is broadly balanced for this model.

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

Buy headroom, not only minimum fit

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

Performance by workload

Workload	Grade	Fit	Decode	TTFT	Context
Chat	S	Tight fit	89.1 tok/s	1186 ms	169K
Coding	S	Tight fit	89.1 tok/s	2174 ms	169K
Agentic Coding	S	Tight fit	89.1 tok/s	3162 ms	169K
Reasoning	S	Tight fit	89.1 tok/s	2569 ms	169K
RAG	S	Tight fit	89.1 tok/s	3952 ms	169K

Inference speed

DeepSeek V4 Flash inference speed — tokens per second by GPU & Mac

Estimated decode speed (tokens/sec) for DeepSeek V4 Flash at NVFP4 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 ~18 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	NVFP4	17.8	Offloads
Mac Studio M2 Ultra 128GB	128 GB	NVFP4	7.0	Too big
Mac Studio M1 Ultra 128GB	128 GB	NVFP4	6.7	Too big
MacBook Pro M4 Max 128GB	128 GB	NVFP4	5.2	Too big
MacBook Pro M4 Max 64GB	64 GB	NVFP4	5.2	Too big
RTX 5090 32GB	32 GB	NVFP4	3.9	Too big
2× RX 7900 XTX 24GB	48 GB	NVFP4	3.8	Too big
MacBook Pro M3 Max 64GB	64 GB	NVFP4	3.6	Too big
MacBook Pro M1 Max 64GB	64 GB	NVFP4	3.3	Too big
MacBook Pro M4 Pro 48GB	48 GB	NVFP4	3.2	Too big
2× RTX 4090 24GB	48 GB	NVFP4	2.6	Too big
RTX 4090 24GB	24 GB	NVFP4	2.5	Too big
2× RTX 3090 24GB	48 GB	NVFP4	2.3	Too big
RX 7900 XTX 24GB	24 GB	NVFP4	2.2	Too big
RTX 3090 24GB	24 GB	NVFP4	2.1	Too big
RTX 4080 Super 16GB	16 GB	NVFP4	2.0	Too big
RTX 4070 12GB	12 GB	NVFP4	2.0	Too big
RTX 3060 12GB	12 GB	NVFP4	2.0	Too big
RTX 4060 8GB	8 GB	NVFP4	2.0	Too big
4× RTX 3060 12GB	48 GB	NVFP4	2.0	Too big

Estimates for single-stream decoding at NVFP4; 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 DeepSeek V4 Flash (284B params) fits at each quantization level on AMD Instinct MI300X 192GB (192.0 GB usable).

Quant	Bits	VRAM	Quality	Fit
Q2_K	2	110.8 GB	Low	S90
Q3_K_SBest for your GPU	3	139.2 GB	Low	S90
NVFP4	4	159.0 GB	Medium	F0
Q4_K_M	4	173.2 GB	Medium	F0
Q5_K_M	5	204.5 GB	High	F0
Q6_K	6	232.9 GB	High	F0
Q8_0	8	303.9 GB	Very High	F0
F16	16	582.2 GB	Maximum	F0

Get started

Copy-paste commands to run DeepSeek V4 Flash on your machine.

Run

docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
  --hf-repo "deepseek-ai/DeepSeek-V4-Flash" \
  --hf-file "DeepSeek-V4-Flash-NVFP4.gguf" \
  -c 4096 -ngl 99

Frequently asked questions

Can AMD Instinct MI300X 192GB run DeepSeek V4 Flash?

Yes, AMD Instinct MI300X 192GB can run DeepSeek V4 Flash with a S grade (Tight fit). Expected decode speed: 89.1 tok/s.

How much VRAM does DeepSeek V4 Flash need?

DeepSeek V4 Flash (284B parameters) requires approximately 179.4 GB of memory with NVFP4 quantization.

What is the best quantization for DeepSeek V4 Flash?

The recommended quantization for DeepSeek V4 Flash is NVFP4, which balances quality and memory efficiency.

What speed will DeepSeek V4 Flash run at on AMD Instinct MI300X 192GB?

On AMD Instinct MI300X 192GB, DeepSeek V4 Flash achieves approximately 89.1 tokens per second decode speed with a time-to-first-token of 2174ms using NVFP4 quantization.

Can AMD Instinct MI300X 192GB run DeepSeek V4 Flash for coding?

For coding workloads, DeepSeek V4 Flash on AMD Instinct MI300X 192GB receives a S grade with 89.1 tok/s and 169K context.

What context window can DeepSeek V4 Flash use on AMD Instinct MI300X 192GB?

On AMD Instinct MI300X 192GB, DeepSeek V4 Flash can safely use up to 169K tokens of context. The model's official context limit is 1.0M, but available memory constrains the safe maximum.

What should I upgrade first if DeepSeek V4 Flash feels slow on AMD Instinct MI300X 192GB?

Buy headroom, not only minimum fit. A slightly larger memory tier gives you safer context growth and makes the recommendation more future-proof.

See all results for AMD Instinct MI300X 192GB See all hardware for DeepSeek V4 Flash

Embed this result▼

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

<iframe src="https://willitrunai.com/embed/deepseek-v4-flash-on-instinct-mi300x-192gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>

Preview: