48 GB VRAM (+8)
B
Makes the model fit on the accelerator instead of staying completely out of reach.8.9 tok/s decode
Makes the model fit on the accelerator instead of staying completely out of reach.
~$4,650 MSRP
Can DeepSeek R1 Distill 70B run on NVIDIA A100 40GB?
YES — With Q3_K_S
B66Good
○Estimated from fit model
DeepSeek R1 Distill 70B needs ~44.1 GB VRAM. NVIDIA A100 40GB has 40.0 GB. With Q3_K_S quantization, expect ~24 tok/s.
Runtime: llama.cppCapacity: OffloadBandwidth: HighStack: StandardBottleneck: Host offload
Share:
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 R1 Distill 70B at Q4_K_M needs 52.5 GB — too much for NVIDIA A100 40GB (40.0 GB). Runs at Q3_K_S (44.1 GB) with low quality. 2 quantization levels fit.
Select quantization to explore
Q4_K_M (Medium quality) — 52.5 GB, exceeds 40.0 GB available
52.5 GB required40.0 GB available
131% VRAM needed
12.5 GB over capacity — needs offload or smaller quantization
Fit status
Too heavy
Decode
14.1 tok/s
TTFT
13745 ms
Safe context
4K
Memory
52.5 GB / 40.0 GB
Offload
20%
Memory breakdown
Weights42.7 GB
KV Cache4.9 GB
Runtime0.9 GB
Headroom4.0 GB
See how fast it feels
With memory offload — actual speed may be lowerSee how fast it feelsDeepSeek R1 Distill 70B on NVIDIA A100 40GB
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: 14.1 tok/s decode · 13.7s TTFT (warm) · 35 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 3.2 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 | 15.6 tok/s | 6782 ms | 4K |
| Coding | F | Too heavy | 14.1 tok/s | 13745 ms | 4K |
| Agentic Coding | F | Too heavy | 11.7 tok/s | 24109 ms | 4K |
| Reasoning | F | Too heavy | 14.1 tok/s | 16244 ms | 4K |
| RAG | F | Too heavy | 11.7 tok/s | 30137 ms | 4K |
Quantization options
How DeepSeek R1 Distill 70B (70B params) fits at each quantization level on NVIDIA A100 40GB (40.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_KBest for your GPU | 2 | 27.3 GB | Low | A74 |
Q3_K_S | 3 | 34.3 GB | Low | F0 |
NVFP4 | 4 | 39.2 GB | Medium | F0 |
Q4_K_M | 4 | 42.7 GB | Medium | F0 |
Q5_K_M | 5 | 50.4 GB | High | F0 |
Q6_K | 6 | 57.4 GB | High | F0 |
Q8_0 | 8 | 74.9 GB | Very High | F0 |
F16 | 16 | 143.5 GB | Maximum | F0 |
Get started
Copy-paste commands to run DeepSeek R1 Distill 70B on your machine.
Run
ollama run deepseek-r1:70bUpgrade options
Hardware that runs DeepSeek R1 Distill 70B well
48 GB VRAM (+8)
B
Makes the model fit on the accelerator instead of staying completely out of reach.17.8 tok/s decode
Makes the model fit on the accelerator instead of staying completely out of reach.
Raises estimated decode speed by about 26%.
~$4,999 MSRP
64 GB VRAM (+24)
A
Makes the model fit on the accelerator instead of staying completely out of reach.11.9 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.
~$6,500 MSRP
80 GB VRAM (+40)3350 GB/s (+1795)
A
Makes the model fit on the accelerator instead of staying completely out of reach.71.7 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.
~$40,000 MSRP