A pure fp16 cast of the canonical BAAI/bge-m3 weights, in safetensors format. BAAI publishes bge-m3 only as an fp32 pytorch_model.bin (a PyTorch pickle); this repo exists so that Rust/candle consumers (and anyone else) can load the model via memory-mapped safetensors at half the size, with no pickle parsing and no fp32→f16 conversion transient at load time.

Nothing here is fine-tuned, re-trained, quantized, or otherwise altered: every tensor is the bitwise result of tensor.to(torch.float16) applied to the canonical fp32 weights. config.json and tokenizer.json are byte-identical copies from the source revision.

Measured (2026-08-22, candle 0.11, Apple M3 Pro Metal and RTX 4070 SUPER CUDA, independently): f16 inference costs nothing measurable in embedding quality versus fp32 — median pairwise cosine agreement with the fp32/quantized reference implementations differs only in the fifth decimal place (0.999720 vs 0.999721 over an 82-text multilingual corpus, 32 B–8 KiB inputs), while running ~27% faster on Metal with half the resident memory.

MIT, inherited from BAAI/bge-m3. All credit for the model belongs to BAAI / the FlagEmbedding team.