Base model and variant
Meta’s Llama 3.3 70B Instruct is a dense 70.6-billion-parameter text model that lists eight languages. Its repository on Hugging Face is gated: you must accept the license and be approved before downloading. Because of that we read its architecture from the abliterated copy’s configuration file, which is unchanged from the base: 80 layers, 64 attention heads, 8 key-value heads with a head dimension of 128, and a configured context of 131,072 tokens.
huihui-ai’s abliterated version applies the remove-refusals-with-transformers approach, a form of abliteration that edits the weights to suppress a refusal direction. The card does not say which layers were edited, publishes no measurements, and suggests retrying a question if the result is not what you want. It is open to download, unlike the base.
Memory at each precision
A 70B dense model is large for local use. Using llama.cpp’s measured bits per weight, the weights alone take about 40 GiB at Q4_K_M, 70 GiB at Q8_0 and 131 GiB at 16-bit. The KV cache adds 320 KiB per token at 16-bit precision: 10 GiB for 32,768 tokens, 40 GiB at the full 131,072. With 10% overhead, 4-bit weights and a 4,096 to 8,192-token context come to roughly 45 to 47 GiB, which is beyond any single 24 GB GPU: it means two such cards, one 48 GB card, or a machine with large unified memory. The table below and the VRAM estimator show the arithmetic.
bartowski’s GGUF repo provides imatrix quantizations from 2-bit to 8-bit, so you can trade quality for memory without converting the weights yourself.
Hosted options
Featherless serves huihui-ai/Llama-3.3-70B-Instruct-abliterated by name at the same per-token rate it charges for the standard Llama 3.3 70B, with a listed context of 32,768 tokens. MolochAI lists llama-3.3-70b-instruct-abliterated with a 200,704-token context at its flat rate; it does not name the weights, though its id matches this variant’s name. The standard model is widely available and cheaper on routers such as OpenRouter, but it is not refusal-reduced.
The license question
Llama 3.3 is not under an open-source license. The Llama 3.3 Community License allows commercial use with conditions, including attribution (“Built with Llama”), a naming rule for derived models, and a separate license requirement for very large services. It also incorporates an acceptable use policy, whose item 1.h prohibits acting “to intentionally circumvent or remove usage restrictions or other safety measures”.
Whether that clause covers downloading, using or serving an abliterated copy is a legal question. Atlas does not answer it, and nobody should treat a model’s availability on Hugging Face or from a provider as a license to use it. If you plan commercial use, read the license and the policy and take legal advice. Our license comparison sets out the main terms side by side.
Ordered as catalogued. Method, layers and claims are as each creator states them; Atlas has not tested any of these variants.
Estimate for Llama 3.3 70B Instruct with 32,768 tokens of context, one sequence, FP16 KV cache and 10% overhead. Variants of the same architecture need about the same.