Phi-3.5-mini-instruct VRAM requirements

Phi-3.5-mini-instruct needs about 6 GiB of GPU memory at Q4_K_M with an 8K context. 8 GB GPUs like the RTX 4060 are enough. At 32K context it needs about 15.6 GiB, which fits 20 GB cards.

These are estimates from the model's config and file sizes, not measurements. How we calculate

VRAM by quantization

QuantFile sizeTotal at 8KTotal at 32K
BF16 safetensors7.6 GB10.7 GiB–12.7 GiB19.7 GiB–22.0 GiB
GGUF Q8_0 (bartowski)4.1 GB6.9 GiB–7.6 GiB15.9 GiB–17.2 GiB
GGUF Q6_K (bartowski)3.1 GB6.0 GiB–6.7 GiB15.0 GiB–16.3 GiB
GGUF Q5_K_M (bartowski)2.8 GB5.7 GiB–6.4 GiB14.7 GiB–16.0 GiB
GGUF Q4_K_M (bartowski)2.4 GB5.4 GiB–6.0 GiB14.4 GiB–15.6 GiB

Names like Q4_K_M and Q8_0 are quantization levels: fewer bits per weight means a smaller file and a small loss of quality. Totals add the KV cache (memory that grows with context length) and runtime buffers, for one chat at a time. Download files

Which GPUs can run it

MemoryUsableExamplesAt 8KAt 32K
8 GB7.0 GiBRTX 4060, RTX 3060, RTX 5060 Ti +1 moreFitsDoes not fit
10 GB9.0 GiBRTX 3080Runs comfortablyDoes not fit
12 GB11.0 GiBRTX 4070, RTX 3060, RTX 5070 +2 moreRuns comfortablyDoes not fit
16 GB14.0 GiBRTX 4080 SUPER, RTX 5080, RX 7800 XT +5 moreRuns comfortablyDoes not fit
16 GB15.0 GiBRTX 4060 Ti, RTX 4070 Ti SUPERRuns comfortablyTight fit
20 GB18.0 GiBRX 7900 XTRuns comfortablyFits
24 GB22.0 GiBRTX 4090, RX 7900 XTX, RTX 3090Runs comfortablyFits
32 GB30.0 GiBRTX 5090Runs comfortablyRuns comfortably
48 GB46.0 GiBL40SRuns comfortablyRuns comfortably
80 GB78.0 GiBA100, H100Runs comfortablyRuns comfortably
64 GB · unified54.0 GiBApple M2 MaxRuns comfortablyRuns comfortably
128 GB · unified96.0 GiBRyzen AI Max+ 395Runs comfortablyRuns comfortably
128 GB · unified112.0 GiBDGX Spark (GB10)Runs comfortablyRuns comfortably
128 GB · unified116.0 GiBApple M4 MaxRuns comfortablyRuns comfortably

GGUF Q4_K_M (bartowski), one chat at a time. Usable memory is the total minus a default reserve for the OS and display.

Detailed checks

Architecture

The KV cache grows with context: about 3 GiB at 8K and 12 GiB at 32K in FP16.

FAQ

Is the download size the same as VRAM use?

No. The Q4_K_M file is about 2.2 GiB, but with an 8K context the total is about 6 GiB, because the KV cache and runtime buffers sit on top of the weights.

How much memory does a longer context add?

For Phi-3.5-mini-instruct the KV cache is about 3 GiB at 8K tokens and 12 GiB at 32K in FP16. The weights stay the same size.

Which quantization should I use?

Start with Q4_K_M (about 6 GiB at 8K). If you have the memory, BF16 needs about 12.7 GiB and stays closer to the original weights.

Download files

QuantFilesFile sizeEngine
BF16 safetensorsmodel-00001-of-00002.safetensors, model-00002-of-00002.safetensors7.64 GB
7,642,159,104 bytes
vllm, sglang
GGUF Q8_0 (bartowski)Phi-3.5-mini-instruct-Q8_0.gguf4.06 GB
4,061,222,688 bytes
llama.cpp, ollama
GGUF Q6_K (bartowski)Phi-3.5-mini-instruct-Q6_K.gguf3.14 GB
3,135,853,344 bytes
llama.cpp, ollama
GGUF Q5_K_M (bartowski)Phi-3.5-mini-instruct-Q5_K_M.gguf2.82 GB
2,815,276,320 bytes
llama.cpp, ollama
GGUF Q4_K_M (bartowski)Phi-3.5-mini-instruct-Q4_K_M.gguf2.39 GB
2,393,232,672 bytes
llama.cpp, ollama
How this number is calculated

Total = weight file + KV cache + runtime buffers. The range spans low to high runtime overhead for the engine shown.

GGUF Q4_K_M (bartowski) · 8,192 tokens · llama.cpp

Weights (file)
2.23 GiB (2,393,232,672 bytes)
KV cache
3.00 GiB (3,221,225,472 bytes)
Runtime buffers
0.13 GiB–0.75 GiB
Total
5.35 GiB–5.98 GiB
  • Requested context is within the recorded native window of 131072 tokens.
  • Full-resident weight estimates use total parameters, including inactive MoE experts.
  • Weight bytes come from the selected weight files only, not from every file in the repository.
  • Download file size is not peak GPU memory. Runtime layout, allocator padding, and KV are extra.
  • Counted files: Phi-3.5-mini-instruct-Q4_K_M.gguf.
  • Context budget is the number of cached tokens per sequence (prompt plus reserved generation).
  • Concurrent sequences share one model replica; weights are not multiplied by sequence count.
  • KV dtype is independent of weight quantization unless you change it.

How we calculate

Open the calculator with this model

Related pages

Sources

Calculator

Loading catalog snapshot…

Context length

Longer context needs more memory for the KV cache, the model's working memory for the conversation.

More options
Quantization
Runtime

Engine and KV cache type for this estimate.

Workload

Overhead scenario and the prefill workspace.

Topology

How the model is split across devices. Memory is not added together.

Replica count is an advanced deployment option. Quantity is nodes or GPUs, not replicas.

Budget

System reserve and safety margin. An empty number keeps the previous value.

GiB (2^30 bytes)

%

Custom model

Import a config or type the architecture. Nothing is scanned until you import.

Parsed in the browser. The file is not uploaded.

Only huggingface.co, config.json, no token. Private repos are rejected.

Pick a model to see how much memory it needs.