[{"data":1,"prerenderedAt":41},["ShallowReactive",2],{"profile-llama-3-3-70b-abliterated":3},{"slug":4,"title":5,"seoTitle":6,"description":7,"summary":8,"published":9,"updated":9,"reviewed":9,"html":10,"toc":11,"sources":25},"llama-3-3-70b-abliterated","Llama 3.3 70B abliterated","Llama 3.3 70B Abliterated: Weights, License and Hosting","Llama 3.3 70B Instruct abliterated by huihui-ai: base vs variant, GGUF quantizations, VRAM needs at each precision, hosted options and the license questions.","The abliterated Llama 3.3 70B Instruct from huihui-ai is a weight-edited copy of Meta's 70B dense model, published in December 2024 with GGUF files by bartowski. It keeps the base architecture: 80 layers, grouped-query attention and a 131,072-token configured context. The Llama 3.3 license and acceptable use policy still apply, and the policy contains a clause about removing safety measures.","2026-11-14","\u003Ch2 id=\"base-model-and-variant\">Base model and variant\u003C\u002Fh2>\n\u003Cp>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.\u003C\u002Fp>\n\u003Cp>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.\u003C\u002Fp>\n\u003Ch2 id=\"memory-at-each-precision\">Memory at each precision\u003C\u002Fh2>\n\u003Cp>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 \u003Ca href=\"\u002Ftools\u002Fvram-estimator\u002F\">VRAM estimator\u003C\u002Fa> show the arithmetic.\u003C\u002Fp>\n\u003Cp>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.\u003C\u002Fp>\n\u003Ch2 id=\"hosted-options\">Hosted options\u003C\u002Fh2>\n\u003Cp>Featherless serves \u003Ccode>huihui-ai\u002FLlama-3.3-70B-Instruct-abliterated\u003C\u002Fcode> 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 \u003Ccode>llama-3.3-70b-instruct-abliterated\u003C\u002Fcode> 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.\u003C\u002Fp>\n\u003Ch2 id=\"the-license-question\">The license question\u003C\u002Fh2>\n\u003Cp>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 \u003Ca href=\"https:\u002F\u002Fwww.llama.com\u002Fllama3_3\u002Fuse-policy\u002F\" rel=\"noopener\" class=\"ext\">acceptable use policy\u003C\u002Fa>, whose item 1.h prohibits acting “to intentionally circumvent or remove usage restrictions or other safety measures”.\u003C\u002Fp>\n\u003Cp>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 \u003Ca href=\"\u002Flearn\u002Fllm-model-licenses\u002F\">license comparison\u003C\u002Fa> sets out the main terms side by side.\u003C\u002Fp>\n",[12,16,19,22],{"id":13,"text":14,"level":15},"base-model-and-variant","Base model and variant",2,{"id":17,"text":18,"level":15},"memory-at-each-precision","Memory at each precision",{"id":20,"text":21,"level":15},"hosted-options","Hosted options",{"id":23,"text":24,"level":15},"the-license-question","The license question",[26,30,34,38],{"title":27,"publisher":28,"url":29},"huihui-ai\u002FLlama-3.3-70B-Instruct-abliterated","Hugging Face (huihui-ai)","https:\u002F\u002Fhuggingface.co\u002Fhuihui-ai\u002FLlama-3.3-70B-Instruct-abliterated",{"title":31,"publisher":32,"url":33},"bartowski\u002FLlama-3.3-70B-Instruct-abliterated-GGUF","Hugging Face (bartowski)","https:\u002F\u002Fhuggingface.co\u002Fbartowski\u002FLlama-3.3-70B-Instruct-abliterated-GGUF",{"title":35,"publisher":36,"url":37},"Llama 3.3 Community License Agreement","Meta","https:\u002F\u002Fwww.llama.com\u002Fllama3_3\u002Flicense\u002F",{"title":39,"publisher":36,"url":40},"Llama 3.3 Acceptable Use Policy","https:\u002F\u002Fwww.llama.com\u002Fllama3_3\u002Fuse-policy\u002F",1794622336924]