voyage-4-nano

Model design

The model uses query and document prompts, mean pooling and normalized output. Its card advertises 32,000-token inputs and a shared vector space with Voyage 4. The architecture uses Qwen-based custom code; the card does not identify a precise upstream pretrained checkpoint.voyage-4-nano downloadable model card

Checkpoint and documentation

The downloadable configuration specifies 2,048 output labels and Matryoshka sizes of 256, 512, 1,024 and 2,048. The pooling configuration also specifies 2,048 dimensions. These files support the checkpoint dimensions given here.voyage-4-nano checkpoint configurationvoyage-4-nano pooling configuration

MongoDB's overview instead lists a 512-dimensional default with 128 and 256 options.MongoDB Voyage AI model specifications

The Voyage text overview lists a 1,024-dimensional default.Voyage AI text embedding catalog and earlier models

Both overview values conflict with the downloadable checkpoint. This page uses the checkpoint configuration for the weight specifications.voyage-4-nano checkpoint configuration

Official GGUF formats

The official repository provides F16 and Q8_0 formats of the same checkpoint. Its instructions require Unicode NFC normalization in llama.cpp. Earlier community conversions could omit bidirectional attention and the final projection, producing incompatible embeddings despite loading without errors. The repository documents minimum runtime versions.Official voyage-4-nano GGUF formats and runtime notes

Practical use

Practical guidance. Local inference gives control over where input text is processed. Preserve the supplied prompts, pooling and normalization when changing runtimes. A model that loads successfully can still produce incompatible vectors if the conversion omits a projection or changes attention.