Nemotron-3-Embed-8B — Community MLX 4-bit

Unofficial community quantization — not an NVIDIA release.

4-bit MLX build of nvidia/Nemotron-3-Embed-8B-BF16 (revision 8ca3ff38), the top-ranked open embedding model on RTEB at time of writing, quantized for Apple-Silicon Macs. All credit for the base model and its training belongs to NVIDIA; this repo only changes the weight storage format. 4.7 GB on disk; runs in the memory budget of a 16–24 GB machine. Embedding cosine fidelity vs the BF16 reference is 0.991, and retrieval scores on our regression subset sit within 0.004 nDCG@10 of BF16.

The Hugging Face file widget reports "~1B params, BF16 · U32". That is a counting artifact: MLX packs eight 4-bit weights into each uint32 element, so the scanner sees one eighth of the logical weights. The model is the full 8B parameters of the base; BF16 tensors are the unquantized embeddings and per-group quantization scales.

The base model is a Ministral3 encoder with bidirectional attention and mean pooling. Stock mlx-lm runs causal attention and would produce wrong embeddings while appearing to work, so this repo ships a small standalone module (nemotron3_embed_mlx.py) that implements the encoder faithfully. Our fixture suite includes a suffix-sensitivity probe confirming bidirectional attention is active in this build.

Use

import importlib.util
from huggingface_hub import snapshot_download

path = snapshot_download("shadowrock-io/Nemotron-3-Embed-8B-Community-MLX-4bit")
spec = importlib.util.spec_from_file_location("nemo_mlx", f"{path}/nemotron3_embed_mlx.py")
mod = importlib.util.module_from_spec(spec); spec.loader.exec_module(mod)

model, tok = mod.load(path)
docs = mod.encode(model, tok, ["passage: MLX runs on Apple Silicon."])
qry  = mod.encode(model, tok, ["query: what does MLX run on?"])
print((docs @ qry.T))  # cosine similarity (outputs are L2-normalized)

Prefixes matter: prepend query: to queries and passage: to documents, matching the base model card. Nothing adds them for you.

Embeddings are 4096-dim, L2-normalized, mean-pooled. Matryoshka truncation to 2048 or 1024 dims works as in the base model: slice, then re-normalize.

Why group size 32

Local sweeps on an Apple M5 Pro (24 GB), scored as cosine fidelity against the BF16 reference on fixed token-ID-locked fixtures:

Recipe	Fidelity (mean)	Fidelity (min)	Size
4-bit, gs32 (this repo)	0.9912	0.9901	4.7 GB
4-bit, gs64	0.9884	0.9868	4.5 GB
6-bit, gs64	0.9991	0.9990	6.5 GB

6-bit is near-lossless if you have the memory headroom; the conversion script in the source repo reproduces it with --bits 6. We ship gs32 as the best fidelity-per-GB at 4-bit.

MTEB retrieval regression vs BF16

MTEB (v2, mteb 2.18.12) on the laptop subset of our frozen regression suite. BF16 baseline computed with the same harness, adapter, and prefixes on an A100:

Task (nDCG@10)	BF16	MLX 4-bit gs32	Delta
NFCorpus	0.4237	0.4199	−0.0038
SciFact	0.8330	0.8338	+0.0008

Gate: per-task loss ≤ 0.01. Both pass. Raw mteb result JSON lives in this repo under results/; rerun with mteb 2.18.12 and the pinned harness from the source repo to verify.

Pooling fixtures (batch-vs-single, batch order, padding invariance, unit norm, prefix discrimination, suffix sensitivity): all pass; cosines ≥ 0.99994 on invariance checks.

RTEB benchmarks vs NVIDIA's published numbers

The base model's public identity is its RTEB leaderboard ranking (Finance #1, German #1, Code #2, Healthcare #4, Legal #4). We ran the open RTEB datasets from those domains on this 4-bit MLX build; the comparison column is NVIDIA's official per-task result from the mteb results repo, not our reproduction.

Task (nDCG@10)	NVIDIA official BF16	MLX 4-bit (this repo)	Delta
HumanEvalRetrieval	1.0000	1.0000	±0.0000
MBPPRetrieval	0.9560	0.9564	+0.0004
WikiSQLRetrieval	0.9950	0.9946	−0.0004
DS1000Retrieval	0.7646	0.7627	−0.0019
FinanceBenchRetrieval	0.9526	0.9405	−0.0121*
HC3FinanceRetrieval	0.7981	0.7976	−0.0005
FinQARetrieval	0.8871	0.8857	−0.0014
LegalQuAD (German)	0.7718	0.7624	−0.0094
LegalSummarization	0.7666	0.7617	−0.0050
ChatDoctorRetrieval	0.7690	0.7692	+0.0002

* FinanceBench has 150 queries — one ranking flip moves the score ~0.007. The CUDA quants score at or above official on this task (FP8 +0.0006, NVFP4 +0.0019), so we read this as small-task noise rather than systematic 4-bit loss, and report it as measured. Mean delta −0.0030 across the 10 tasks — the same 10-task set appears in the FP8 and NVFP4 cards for direct comparison (those builds add two AILA legal tasks not run here). The private RTEB datasets can only be run by the MTEB team, so this table covers the open subset.

Performance

Test device: MacBook Pro, Apple M5 Pro — 15-core CPU (5 super / 10 performance), 16-core GPU, 24 GB unified memory, macOS 27. The model needs ~5 GB of that budget, so 16 GB machines fit with room for a working set.

Batch size 8, real token counts:

Input length	Texts/s	Tokens/s	p50 batch latency
~28 tok (short query)	39.0	1,092	0.21 s
~102 tok	13.5	1,372	0.60 s
~512 tok	3.5	1,392	2.3 s
~2,048 tok	0.8	1,271	9.9 s

Throughput saturates near 1,300–1,400 tokens/s at document lengths (memory-bandwidth-bound); larger batches add latency, not throughput. For mixed-length corpora, sort by length before batching: padding to the longest batch member dominates cost otherwise.

Quantization details

Method: mlx.nn.quantize, affine, 4-bit, group size 32, applied uniformly to all linear layers; embeddings and norms untouched. No calibration data (data-free quantization).
Converted with scripts/convert_8b.py (in this repo), from base revision 8ca3ff382cf1de715e05acac8b553e0a084680d0.
Raw eval JSON ships under results/; the fixture/eval harness ships under scripts/.

Caveats

The regression subset is English retrieval; the multilingual and long-document tasks in the full suite run on the CUDA artifacts, not this one. Expect the base model's multilingual behavior with 4-bit noise on top, unmeasured here.
mlx-embeddings/mlx-lm do not load this architecture correctly (causal attention). Use the bundled module.
Quantization was validated on macOS 27 / mlx 0.29; older mlx releases may not support the quantized layout.

Intended use & limitations

Intended uses are the base model's: dense retrieval, semantic search, and RAG indexing over text corpora, with query: /passage: prefixed inputs. The base card's intended-use, safety, and language-coverage statements — nvidia/Nemotron-3-Embed-8B-BF16 — carry over unchanged; quantization alters none of the model's behavior boundaries, only its numeric precision. Our evaluation establishes parity on the benchmarks listed above and nothing beyond them: other languages, domains, sequence-length regimes, and hardware/runtime combinations inherit the base model's behavior with quantization noise that we have not measured there.

Attribution & citation

Quantization, validation harness, and card by Matt Busi (@mattbusi on Hugging Face) at ShadowRock. If you use this build, cite the NVIDIA base model — the embedding quality is theirs:

@misc{nvidia2026nemotron3embed,
  title  = {Nemotron-3-Embed-8B},
  author = {NVIDIA},
  year   = {2026},
  url    = {https://huggingface.co/nvidia/Nemotron-3-Embed-8B-BF16}
}

License

OpenMDW-1.1, inherited from the base model (see LICENSE). NOTICE carries the upstream Apache-2.0 attribution for the Ministral component plus our modification statement. Community build by ShadowRock; no NVIDIA affiliation or endorsement.

About ShadowRock

ShadowRock is an AI-specialized systems integrator and Zendesk Premier Partner. We help businesses get real value from their go-to-market technology, from CRM and support platforms to applied AI like the models in this collection. Find us at shadowrock.io or on LinkedIn.