Llama-3.3-70B-Instruct GGUF size and VRAM requirements

License: llama3.3 ⬇ 20,988 ❤ 75
Parameters70.55B
Context131,072

Llama-3.3-70B-Instruct is a 70.55 billion parameter AI model from the Llama family, designed for instruction-following tasks and licensed under the llama3.3 terms. It is optimized for general-purpose reasoning and interaction, with knowledge current as of December 2023. The model is compatible with tools like LM Studio and can be accessed via Hugging Face.

Under the hood it uses 80 transformer layers, a hidden size of 8,192, 64 attention heads. It uses grouped-query attention (64 query heads sharing 8 key/value heads), which already trims KV-cache memory compared with full multi-head attention.

To run bartowski/Llama-3.3-70B-Instruct-GGUF locally at a 4,096-token context, its quantized versions need between 17.65 GB (IQ1_M, lowest quality) and 133.48 GB (F16, highest quality) of memory, weights plus KV cache and a system margin included. Context length drives that memory directly: at 4,096 tokens the KV cache for Llama-3.3-70B-Instruct-GGUF is about 1.25 GB, rising to roughly 40.0 GB at its full 131,072-token context. Shorter prompts free up memory for a higher-quality quantization.

For most users the best balance is IQ2_S, needing about 22.76 GB. That means bartowski/Llama-3.3-70B-Instruct-GGUF fits entirely in the VRAM of a 24 GB GPU or larger, running fully on the GPU.

Available GGUF quantizations for bartowski/Llama-3.3-70B-Instruct-GGUF include IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, Q2_K_L, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, Q3_K_XL, Q4_0_4_4, Q4_0_4_8, Q4_0_8_8, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_K_L, Q5_K_S, Q5_K_M, Q5_K_L, Q6_K, Q6_K_L, Q8_0, F16. The model supports a native context length of up to 131,072 tokens; a longer context grows the KV cache and the memory needed.

→ Guide: How much VRAM do you need?

GGUF file size and memory by quantization

Compare real GGUF weight sizes, estimated KV cache and total memory for Q4, Q5, Q8 and every quantization published in this repository.

Quant.Bits QualityWeights KVTotal Speed~Verdict
IQ1_M 1.9 Very low 15.6 GB 1.25 GB 17.65 GB 3.2 t/s Offload
IQ2_XXS 2.17 Very low 17.79 GB 1.25 GB 19.84 GB 2.8 t/s Offload
IQ2_XS 2.4 Very low 19.69 GB 1.25 GB 21.74 GB 2.5 t/s Offload
IQ2_S 2.52 Very low 20.71 GB 1.25 GB 22.76 GB 2.4 t/s Offload
IQ2_M 2.73 Low 22.46 GB 1.25 GB 24.51 GB Insufficient
Q2_K 2.99 Low 24.56 GB 1.25 GB 26.61 GB Insufficient
Q2_K_L 3.11 Low 25.52 GB 1.25 GB 27.57 GB Insufficient
IQ3_XXS 3.11 Low 25.58 GB 1.25 GB 27.63 GB Insufficient
IQ3_XS 3.32 Fair 27.29 GB 1.25 GB 29.34 GB Insufficient
Q3_K_S 3.51 Fair 28.79 GB 1.25 GB 30.84 GB Insufficient
IQ3_M 3.62 Fair 29.74 GB 1.25 GB 31.79 GB Insufficient
Q3_K_M 3.89 Fair 31.91 GB 1.25 GB 33.96 GB Insufficient
Q3_K_L 4.21 Good 34.59 GB 1.25 GB 36.64 GB Insufficient
IQ4_XS 4.3 Good 35.3 GB 1.25 GB 37.35 GB Insufficient
Q3_K_XL 4.32 Good 35.45 GB 1.25 GB 37.5 GB Insufficient
Q4_0_4_4 4.53 Good 37.22 GB 1.25 GB 39.27 GB Insufficient
Q4_0_4_8 4.53 Good 37.22 GB 1.25 GB 39.27 GB Insufficient
Q4_0_8_8 4.53 Good 37.22 GB 1.25 GB 39.27 GB Insufficient
IQ4_NL 4.54 Good 37.3 GB 1.25 GB 39.35 GB Insufficient
Q4_0 4.55 Good 37.36 GB 1.25 GB 39.41 GB Insufficient
Q4_K_S 4.57 Good 37.58 GB 1.25 GB 39.63 GB Insufficient
Q4_K_M 4.82 Good 39.6 GB 1.25 GB 41.65 GB Insufficient
Q4_K_L 4.91 Good 40.33 GB 1.25 GB 42.38 GB Insufficient
Q5_K_S 5.52 Very good 45.32 GB 1.25 GB 47.37 GB Insufficient
Q5_K_M 5.66 Very good 46.52 GB 1.25 GB 48.57 GB Insufficient
Q5_K_L 5.74 Very good 47.12 GB 1.25 GB 49.17 GB Insufficient
Q6_K 6.56 Excellent 53.91 GB 1.25 GB 55.96 GB Insufficient
Q6_K_L 6.62 Excellent 54.39 GB 1.25 GB 56.44 GB Insufficient
Q8_0 8.5 Excellent 69.83 GB 1.25 GB 71.88 GB Insufficient
F16 16.0 Excellent 131.43 GB 1.25 GB 133.48 GB Insufficient

KV cache computed from the model's exact architecture. Speed is a rough estimate bounded by memory bandwidth.

Frequently asked questions

What kind of model is bartowski/Llama-3.3-70B-Instruct-GGUF?

bartowski/Llama-3.3-70B-Instruct-GGUF is an instruction-tuned chat model with 70.55 billion parameters, based on the llama architecture. It is released under the llama3.3 license and distributed as GGUF files for local inference.

How much VRAM do you need to run bartowski/Llama-3.3-70B-Instruct-GGUF?

You need about 22.76 GB of VRAM to run bartowski/Llama-3.3-70B-Instruct-GGUF entirely on the GPU using the IQ2_S quantization (at a 4,096-token context). Smaller quantizations lower the requirement at the cost of quality.

Can I run bartowski/Llama-3.3-70B-Instruct-GGUF on an 8 GB GPU?

Partially. bartowski/Llama-3.3-70B-Instruct-GGUF only fits on an 8 GB GPU by offloading part of it to system RAM (with IQ2_S), which runs but is slower.

Can I run bartowski/Llama-3.3-70B-Instruct-GGUF on a 16 GB GPU?

Partially. bartowski/Llama-3.3-70B-Instruct-GGUF only fits on a 16 GB GPU by offloading part of it to system RAM (with Q5_K_S), which runs but is slower.

Can I run bartowski/Llama-3.3-70B-Instruct-GGUF on a 24 GB GPU?

Yes. With 24 GB of VRAM you can run bartowski/Llama-3.3-70B-Instruct-GGUF fully on the GPU using IQ2_S (about 22.76 GB).

What context length does bartowski/Llama-3.3-70B-Instruct-GGUF support?

bartowski/Llama-3.3-70B-Instruct-GGUF supports a native context length of up to 131,072 tokens. A longer context grows the KV cache, so it increases the memory needed to run the model.

What is the best quantization for bartowski/Llama-3.3-70B-Instruct-GGUF?

For bartowski/Llama-3.3-70B-Instruct-GGUF, a strong default is Q4_K_M, which needs about 41.65 GB and keeps most of the quality while roughly halving the memory versus 8-bit. With VRAM to spare, Q5_K_M or Q6_K add a little more quality; if you are tight on memory, a smaller quantization still runs. Pick the highest quantization that fits your VRAM.