Is unified memory on Mac Studio M3 Ultra 256GB as fast as VRAM for DeepSeek V2.5 236B?

Not always. Mac Studio M3 Ultra 256GB can often fit larger models thanks to unified memory, but a discrete GPU with dedicated high-bandwidth VRAM may still decode faster once the model fits. For this combination, the important distinction is capacity versus sustained throughput.

Can DeepSeek V2.5 236B run on Mac Studio M3 Ultra 256GB?

Q: What should I upgrade first if DeepSeek V2.5 236B feels slow on Mac Studio M3 Ultra 256GB?

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.

YES — With NVFP4

B66Good

Estimated from fit model

DeepSeek V2.5 236B needs ~219.3 GB VRAM. Mac Studio M3 Ultra 256GB has 184.3 GB. With NVFP4 quantization, expect ~14 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.

DeepSeek V2.5 236B at Q4_K_M needs 231.1 GB — too much for Mac Studio M3 Ultra 256GB (184.3 GB). Runs at NVFP4 (219.3 GB) with medium quality. 3 quantization levels fit.

Capabilities:

Select quantization to explore

Q4_K_M (Medium quality) — 231.1 GB, exceeds 184.3 GB available

231.1 GB required184.3 GB available

125% VRAM needed

46.8 GB over capacity — needs offload or smaller quantization

Fit status

Too heavy

Decode

11.7 tok/s

TTFT

16613 ms

Safe context

Memory

231.1 GB / 184.3 GB

Offload

20%

Memory breakdown

Weights144.0 GB

KV Cache58.6 GB

Runtime0.9 GB

Headroom27.6 GB

See how fast it feels

With memory offload — actual speed may be lower

See how fast it feelsDeepSeek V2.5 236B on Mac Studio M3 Ultra 256GB
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: 11.7 tok/s decode · 16.6s TTFT (warm) · 29 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 20% 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.

Shared-memory contention still exists

The OS, browser, and inference runtime all compete for the same physical memory pool, so real-world headroom is less forgiving than raw capacity suggests.

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 21.1 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	A	Very compromised (needs ~12.5 GB host RAM)	13.9 tok/s	7613 ms	4K
Coding	F	Too heavy	11.7 tok/s	16613 ms	4K
Agentic Coding	F	Too heavy	9.0 tok/s	31425 ms	4K
Reasoning	F	Too heavy	11.7 tok/s	19634 ms	4K
RAG	F	Too heavy	9.0 tok/s	39281 ms	4K

Quantization options

How DeepSeek V2.5 236B (236B params) fits at each quantization level on Mac Studio M3 Ultra 256GB (184.3 GB usable).

Quant	Bits	VRAM	Quality	Fit
Q2_K	2	92.0 GB	Low	A82
Q3_K_S	3	115.6 GB	Low	A82
NVFP4	4

Get started

Copy-paste commands to run DeepSeek V2.5 236B on your machine.

Run

docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
  --hf-repo "deepseek-ai/DeepSeek-V2.5" \
  --hf-file "DeepSeek-V2.5-Q4_K_M.gguf" \
  -c 4096 -ngl 99

Upgrade options

Hardware that runs DeepSeek V2.5 236B well

AMD Instinct MI350X 288GBBudget pick

288 GB VRAM (+32)8000 GB/s (+7181)

Makes the model fit on the accelerator instead of staying completely out of reach.109.3 tok/s decode

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.

~$8,000 MSRP

AMD Instinct MI300X 192GBBest value

5300 GB/s (+4481)

Makes the model fit on the accelerator instead of staying completely out of reach.42.5 tok/s decode

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

Raises estimated decode speed by about 263%.

~$15,000 MSRP

Frequently asked questions

Can Mac Studio M3 Ultra 256GB run DeepSeek V2.5 236B?

Yes, Mac Studio M3 Ultra 256GB can run DeepSeek V2.5 236B at NVFP4 quantization (Very compromised (needs ~21.1 GB host RAM)). The recommended Q4_K_M requires 231.1 GB which exceeds available memory, but at NVFP4 it needs only 219.3 GB. Expected decode speed: 14.2 tok/s.

How much VRAM does DeepSeek V2.5 236B need?

DeepSeek V2.5 236B (236B parameters) requires approximately 231.1 GB at Q4_K_M quantization. On Mac Studio M3 Ultra 256GB, it fits at NVFP4 using 219.3 GB.

What is the best quantization for DeepSeek V2.5 236B?

The recommended quantization is Q4_K_M, but on Mac Studio M3 Ultra 256GB the best fitting quantization is NVFP4, which uses 219.3 GB.

What speed will DeepSeek V2.5 236B run at on Mac Studio M3 Ultra 256GB?

On Mac Studio M3 Ultra 256GB, DeepSeek V2.5 236B achieves approximately 14.2 tokens per second decode speed with a time-to-first-token of 13613ms using NVFP4 quantization.

Can Mac Studio M3 Ultra 256GB run DeepSeek V2.5 236B for coding?

For coding workloads, DeepSeek V2.5 236B on Mac Studio M3 Ultra 256GB receives a F grade with 11.7 tok/s and 4K context.

What context window can DeepSeek V2.5 236B use on Mac Studio M3 Ultra 256GB?

On Mac Studio M3 Ultra 256GB, DeepSeek V2.5 236B can safely use up to 6K tokens of context at NVFP4 quantization. The model's official context limit is 131K, but available memory constrains the safe maximum.

See all results for Mac Studio M3 Ultra 256GB See all hardware for DeepSeek V2.5 236B

Embed this result▼

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

<iframe src="https://willitrunai.com/embed/deepseek-v2.5-236b-on-m3-ultra-256gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>

Preview: