8,030,261,248 (8.03B; dense, including untied input embeddings and output head)
8,192 tokens configured in the original export; validated maximum unknown
Publisher-declared Llama 3.3 Community License Agreement + Acceptable Use Policy; custom restricted license, not Apache/MIT. Redistribution requires the agreement, Built with Llama attribution and NOTICE; distributed derived model names must begin with Llama. More than 700 million release-date monthly active users requires a separate Meta grant; no revenue threshold. Original repository lacks standalone LICENSE/NOTICE files; historical API terms and applicability to the reconstructed checkpoint remain review flags.
Community-published BF16, text-only autoregressive LlamaForCausalLM checkpoint, not a verified direct Meta public-weight release. Meta's April 2025 LlamaCon announcement confirms a new Llama 3.3 8B model for API fine-tuning, but that limited-preview service announcement is not this checkpoint's public-weight date. Allura's publisher reports downloading a merged fine-tuned model plus its adapter from Meta's API and subtracting the adapter to reconstruct the pre-adapter instruction model; the subtraction code, original export and numerical recovery accuracy were not independently verified. Export metadata reportedly names meta-llama/Meta-Llama-3-8B-Instruct, so behavioral differences and benchmark results do not by themselves prove exact Llama 3.3 ancestry or equivalence to the hosted model. The downloaded checkpoint has 32 dense Transformer layers, width 4096, SwiGLU intermediate width 14336, grouped-query attention with 32 query and 8 key/value heads, head dimension 128 and an untied 128256-token vocabulary. Original configuration sets 8192 positions, RoPE theta 500000 and no RoPE scaling; the tokenizer's huge model_max_length is a sentinel, and generation_config max_length 4096 is a generation default rather than the context ceiling. A Llama-header chat template is provided in a separate Jinja file. The publisher's OpenBench evaluation at temperature 1.0 reports IFEval 81.95 versus 78.2 for Llama 3.1 8B, averaging strict/loose instruction/prompt accuracies over one epoch, and GPQA Diamond 37.0 versus 29.3 over three epochs; these are community-reported measurements, not Meta's official benchmark results or independently reproduced scores. The publisher acknowledges factual hallucinations and withdrew tau-bench scores because looping traces and apparently incorrect successes made scoring unreliable. The exact checkpoint's pretraining corpus, token budget, instruction-tuning and reinforcement-learning recipe, distillation history and knowledge cutoff remain undisclosed. December 30 UTC is a provisional available-by date supported by a public download discussion, community conversion and local Q4 execution; those early events fall on December 29 in UTC-05, and exact first public access remains unresolved. This card covers only Allura's original 8K BF16 export. Bartowski's related GGUF packages and shb777's 128K RoPE/template/generation-config repackaging are separately documented; the latter has identical four shard hashes, not demonstrated new training or a verified native 128K Meta export.