OpenGryd/Nemotron-3.5-Embed-1B-BF16-MLX

Nemotron-3.5-Embed-1B-BF16 (MLX)

An MLX build of nvidia/Nemotron-3.5-Embed-1B-BF16 (snapshot e24ca1e3), produced by OpenGryd. The safetensors are the upstream weights, unmodified — MLX loads BF16 safetensors directly, so this repository adds only the serving module (nemotron35_embed_mlx.py) plus the original configs and tokenizer.

What is served

Nemotron-3.5-Embed-1B pairs a Pixtral vision encoder with a bidirectional Ministral3 language tower (2048-dim mean-pooled, L2-normalised embeddings; 32,768-token context). This MLX build serves the text path: the tower alone. Image inputs are not wired here — vision_tower.* / multi_modal_projector.* tensors ship in the safetensors unchanged for any future vision path, and are skipped at load time.

mlx_lm.load / mlx_vlm.load will pull the wrong (causal) variant and produce garbage embeddings — use this repository's own module:

import mlx.core as mx
import nemotron35_embed_mlx as nem

model, tok = nem.load("opengryd/Nemotron-3.5-Embed-1B-BF16-MLX")

texts = ["query: What is NeMo?", "passage: NeMo Framework …"]
ids = tok(texts, padding=True, truncation=True, max_length=32768, return_tensors="np")
emb = model(mx.array(ids["input_ids"]), mx.array(ids["attention_mask"]))
# emb: [B, 2048], fp32, L2-normalised — dot product = cosine similarity

encode(model, tokenizer, texts, input_type="query"|"passage") applies the upstream query: / passage: prefixes for asymmetric retrieval.

Fidelity

The text tower is identically configured to mlx-community/Nemotron-3-Embed-1B-BF16 (whose bidirectional-encoder approach this module follows). Verified against the PyTorch reference (transformers 5.20, CPU fp32, identical right-padded batches, multilingual texts incl. CJK): worst-case cosine 0.9997 across batches (conversion validated 2026-11-12; house bar > 0.999, same standard as the Nemotron-3-Embed-1B MLX conversion).

License

OpenMDW-1.1 (see LICENSE, NOTICE, THIRD_PARTY_NOTICES.md, inherited verbatim from the upstream release). The Ministral base models are Apache-2.0.