Can InternLM 20B run on Mac Studio M3 Ultra 96GB?

YES — Runs Great

B61Good

○Estimated from fit model

InternLM 20B needs ~44.0 GB VRAM. Mac Studio M3 Ultra 96GB has 69.1 GB. With Q5_K_M quantization, expect ~39 tok/s.

Runtime: llama.cppCapacity: RoomyBandwidth: 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

Q5_K_M (High quality) — 44.0 GB, 39.4 tok/s, Runs well

44.0 GB required69.1 GB available

64% VRAM used

Fit status

Runs well

Decode

39.4 tok/s

TTFT

4908 ms

Safe context

Memory

44.0 GB / 69.1 GB

Memory breakdown

Weights14.4 GB

KV Cache18.3 GB

Runtime0.9 GB

Headroom10.4 GB

See how fast it feels

See how fast it feelsInternLM 20B on Mac Studio M3 Ultra 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: 39.4 tok/s decode · 4.9s TTFT (warm) · 99 tok/s prefill

What limits this setup

This setup is broadly balanced for this model.

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

Performance by workload

Workload	Grade	Fit	Decode	TTFT	Context
Chat	B	Runs well	39.4 tok/s	2677 ms	8K
Coding	B	Runs well	39.4 tok/s	4908 ms	8K
Agentic Coding	B	Tight fit	39.4 tok/s	7138 ms	8K
Reasoning	B	Runs well	39.4 tok/s	5800 ms	8K
RAG	B	Tight fit	39.4 tok/s	8923 ms	8K

Quantization options

How InternLM 20B (20B params) fits at each quantization level on Mac Studio M3 Ultra 96GB (69.1 GB usable).

Quant	Bits	VRAM	Quality	Fit
Q2_K	2	7.8 GB	Low	C48
Q3_K_S	3	9.8 GB	Low	C49
NVFP4	4	11.2 GB	Medium	C49
Q4_K_M	4	12.2 GB	Medium	C49
Q5_K_M	5	14.4 GB	High	C50
Q6_K	6	16.4 GB	High	C50
Q8_0	8	21.4 GB	Very High	C51
F16Best for your GPU	16	41.0 GB	Maximum	B56

Get started

Copy-paste commands to run InternLM 20B on your machine.

Run

docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
 --hf-repo "internlm/internlm2_5-20b-chat" \
 --hf-file "internlm2_5-20b-chat-Q5_K_M.gguf" \
 -c 4096 -ngl 99

Upgrade options

Hardware that runs InternLM 20B well

👁 NVIDIA
NVIDIA A100 80GBBudget pick

2039 GB/s (+1220)

Raises estimated decode speed by about 208%.121.3 tok/s decode

Raises estimated decode speed by about 208%.

~$15,000 MSRP

👁 NVIDIA
NVIDIA A800 80GBBest value

1935 GB/s (+1116)

Raises estimated decode speed by about 171%.106.9 tok/s decode

Raises estimated decode speed by about 171%.