VIDRAFT, a Korean Pre-AGI startup, released Darwin-180B-RSI-R3, a 180-billion-parameter open-weight language model distributed in the GGUF binary container format used by the llama.cpp ecosystem. A Chinese technical outlet, ZAKER, conducted a byte-level structural dissection of the GGUF artifact, detailing its internal organization: a magic number and version header for parser compatibility, a typed metadata key-value section storing tokenizer configuration, architecture hyperparameters (layer count, attention heads, context length), and provenance data, followed by sequentially laid-out tensor data blocks with alignment padding, each described by name, shape, and quantization type. GGUF supports quantization schemes such as Q4_K, Q5_K, and Q8_0, enabling memory-precision trade-offs critical for deploying a 180B model on consumer or server hardware where VRAM is insufficient for full-precision weights; its memory-mapped I/O allows runtime loading of only required tensor pages. The RSI-R3 suffix denotes a specific training or alignment revision within the Darwin lineage, though VIDRAFT has not publicly disclosed the precise iterative refinement methodology. The ZAKER analysis provides no benchmark metrics—accuracy, perplexity, or throughput—serving strictly as a format reference. The model is publicly available via VIDRAFT's Hugging Face organization (repository path to be verified on their official profile) and loadable via standard llama.cpp invocation, with inference flags for context offloading and GPU layer splitting recommended for multi-GPU setups. License terms remain unspecified in the source coverage, requiring verification on the model card before commercial deployment.
Read original
dev.to