Can DeepSeek LLM 67B run on RTX 5090 32GB?

YES — With Q2_K

C48Usable

Estimated from fit model

DeepSeek LLM 67B needs ~36.0 GB VRAM. RTX 5090 32GB has 32.0 GB. With Q2_K quantization, expect ~17 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 LLM 67B at Q4_K_M needs 50.8 GB — too much for RTX 5090 32GB (32.0 GB). Runs at Q2_K (36.0 GB) with low quality.

Capabilities:

Select quantization to explore

Q4_K_M (Medium quality) — 50.8 GB, exceeds 32.0 GB available

50.8 GB required32.0 GB available

159% VRAM needed

18.8 GB over capacity — needs offload or smaller quantization

Fit status

Too heavy

Decode

6.2 tok/s

TTFT

31355 ms

Safe context

Memory

50.8 GB / 32.0 GB

Offload

40%

Memory breakdown

Weights40.9 GB

KV Cache5.8 GB

Runtime0.9 GB

Headroom3.2 GB

See how fast it feels

With memory offload — actual speed may be lower

See how fast it feelsDeepSeek LLM 67B on RTX 5090 32GB
1st promptCold start — includes initialization
>Solve: If a train travels at 60 mph for 2.5 hours, how far does it go?
Let me work through this step by step.

Given information:
  - Speed = 60 mph (miles per hour)
  - Time  = 2.5 hours

Formula: Distance = Speed x Time

Calculation:
  Distance = 60 mph x 2.5 h
           = 150 miles

The train travels 150 miles.

Double-check: at 60 mph the train covers 60 miles each hour. In half an hour it covers 30 miles. So in 2.5 hours: 60 + 60 + 30 = 150 miles. Confirmed.
2nd promptWarm — model ready, no init overhead
>What if the train stops for 20 minutes halfway?
The stop doesn't change the total distance covered while moving, but it affects total trip time.

Breakdown:
  - First half: 1.25 hours at 60 mph = 75 miles
  - Stop: 20 minutes (no distance)
  - Second half: 1.25 hours at 60 mph = 75 miles

Total distance: 150 miles (same)
Total time: 2.5 h + 0.33 h = 2.83 hours
Average speed: 150 / 2.83 = 53 mph

The distance stays the same but average speed drops to 53 mph because of the stop.
Estimated: 6.2 tok/s decode · 31.4s TTFT (warm) · 15 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 2.9 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	7.0 tok/s	15169 ms	4K
Coding	F	Too heavy	6.2 tok/s	31355 ms	4K
Agentic Coding	F	Too heavy	5.0 tok/s	56865 ms	4K
Reasoning	F	Too heavy	6.2 tok/s	37056 ms	4K
RAG	F	Too heavy	5.0 tok/s	71081 ms	4K

Quantization options

How DeepSeek LLM 67B (67B params) fits at each quantization level on RTX 5090 32GB (32.0 GB usable).

Quant	Bits	VRAM	Quality	Fit
Q2_K	2	26.1 GB	Low	F0
Q3_K_S	3	32.8 GB	Low	F0
NVFP4	4	37.5 GB	Medium	F0
Q4_K_M	4	40.9 GB	Medium	F0
Q5_K_M	5	48.2 GB	High	F0
Q6_K	6	54.9 GB	High	F0
Q8_0	8	71.7 GB	Very High	F0
F16	16	137.4 GB	Maximum	F0

Get started

Copy-paste commands to run DeepSeek LLM 67B on your machine.

Run

docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
  --hf-repo "deepseek-ai/deepseek-llm-67b-chat" \
  --hf-file "deepseek-llm-67b-chat-Q4_K_M.gguf" \
  -c 4096 -ngl 99

Upgrade options

Hardware that runs DeepSeek LLM 67B well

RTX A6000 48GBBest value

48 GB VRAM (+16)

Makes the model fit on the accelerator instead of staying completely out of reach.9.7 tok/s decode

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

Raises estimated decode speed by about 56%.

~$4,650 MSRP

RTX PRO 5000 Blackwell 48GBNVIDIA upgrade

48 GB VRAM (+16)

Makes the model fit on the accelerator instead of staying completely out of reach.19.2 tok/s decode

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

Raises estimated decode speed by about 210%.

~$4,999 MSRP

NVIDIA A16 64GBBudget pick

64 GB VRAM (+32)

Makes the model fit on the accelerator instead of staying completely out of reach.12.5 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

NVIDIA H100 80GBBiggest leap

80 GB VRAM (+48)3350 GB/s (+1558)

Makes the model fit on the accelerator instead of staying completely out of reach.74.9 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.

~$40,000 MSRP

Frequently asked questions

Can RTX 5090 32GB run DeepSeek LLM 67B?

Yes, RTX 5090 32GB can run DeepSeek LLM 67B at Q2_K quantization (Very compromised (needs ~2.9 GB host RAM)). The recommended Q4_K_M requires 50.8 GB which exceeds available memory, but at Q2_K it needs only 36.0 GB. Expected decode speed: 16.5 tok/s.

How much VRAM does DeepSeek LLM 67B need?

DeepSeek LLM 67B (67B parameters) requires approximately 50.8 GB at Q4_K_M quantization. On RTX 5090 32GB, it fits at Q2_K using 36.0 GB.

What is the best quantization for DeepSeek LLM 67B?

The recommended quantization is Q4_K_M, but on RTX 5090 32GB the best fitting quantization is Q2_K, which uses 36.0 GB.

What speed will DeepSeek LLM 67B run at on RTX 5090 32GB?

On RTX 5090 32GB, DeepSeek LLM 67B achieves approximately 16.5 tokens per second decode speed with a time-to-first-token of 11716ms using Q2_K quantization.

Can RTX 5090 32GB run DeepSeek LLM 67B for coding?

For coding workloads, DeepSeek LLM 67B on RTX 5090 32GB receives a F grade with 6.2 tok/s and 4K context.

What context window can DeepSeek LLM 67B use on RTX 5090 32GB?

On RTX 5090 32GB, DeepSeek LLM 67B can safely use up to 4K tokens of context at Q2_K quantization. The model's official context limit is 4K, but available memory constrains the safe maximum.

What should I upgrade first if DeepSeek LLM 67B feels slow on RTX 5090 32GB?

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 RTX 5090 32GB See all hardware for DeepSeek LLM 67B

Embed this result▼

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

<iframe src="https://willitrunai.com/embed/deepseek-llm-67b-on-rtx-5090-32gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>

Preview:

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 2.9 GB of extra host RAM just for the offloaded portion, before OS and other tools.

Workload

Grade

Fit

Decode

TTFT

Context

Chat

Too heavy

7.0 tok/s

15169 ms

Coding

Too heavy

6.2 tok/s

31355 ms

Agentic Coding

Too heavy

5.0 tok/s

56865 ms

Reasoning

Too heavy

6.2 tok/s

37056 ms

RAG

Too heavy

5.0 tok/s

71081 ms

Quant

Bits

VRAM

Quality

Fit

Q2_K

26.1 GB

Low

Q3_K_S

32.8 GB

Low

NVFP4

37.5 GB

Medium

Q4_K_M

40.9 GB

Medium

Q5_K_M

48.2 GB

High

Q6_K

54.9 GB

High

Q8_0

71.7 GB

Very High

F16

137.4 GB

Maximum

Frequently asked questions

Can RTX 5090 32GB run DeepSeek LLM 67B?

How much VRAM does DeepSeek LLM 67B need?

DeepSeek LLM 67B (67B parameters) requires approximately 50.8 GB at Q4_K_M quantization. On RTX 5090 32GB, it fits at Q2_K using 36.0 GB.

What is the best quantization for DeepSeek LLM 67B?

The recommended quantization is Q4_K_M, but on RTX 5090 32GB the best fitting quantization is Q2_K, which uses 36.0 GB.

What speed will DeepSeek LLM 67B run at on RTX 5090 32GB?

On RTX 5090 32GB, DeepSeek LLM 67B achieves approximately 16.5 tokens per second decode speed with a time-to-first-token of 11716ms using Q2_K quantization.

Can RTX 5090 32GB run DeepSeek LLM 67B for coding?

For coding workloads, DeepSeek LLM 67B on RTX 5090 32GB receives a F grade with 6.2 tok/s and 4K context.

What context window can DeepSeek LLM 67B use on RTX 5090 32GB?

On RTX 5090 32GB, DeepSeek LLM 67B can safely use up to 4K tokens of context at Q2_K quantization. The model's official context limit is 4K, but available memory constrains the safe maximum.

What should I upgrade first if DeepSeek LLM 67B feels slow on RTX 5090 32GB?

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.