Best Embedding Models to Run Locally in 2026

Quick Answer

Qwen’s Qwen3-VL-Embedding-8B (Qwen) is the top pick as of September 2026 solely through the downloads tiebreak, with no public benchmark yet for any candidate to establish a quality advantage (MTEB results, archived date unknown) [1][3][4]. In order, the ranking is Qwen3-VL-Embedding-8B (Qwen), Qwen3-VL-Embedding-2B (Qwen), and Nemotron-3-Embed-1B-BF16 (nvidia) [1][3].

Key Takeaways

• Qwen’s Qwen3-VL-Embedding-8B ranks first because its 1,210,187 downloads over the preceding 30 days lead the popularity tiebreak as of September 25, 2026—not because of a demonstrated benchmark advantage [1][3][4].
• Qwen’s Qwen3-VL-Embedding-2B ranks second with 1,009,536 downloads over the same period; its 2.1B parameters and 4.3 GB of BF16 weights compare with the first-ranked model’s 8.1B parameters and 16.3 GB of BF16 weights [3][1].
• nvidia’s Nemotron-3-Embed-1B-BF16 ranks third with 519,334 downloads over the same period, 1.1B parameters and 2.3 GB of BF16 weights; its July 14, 2026 release satisfies the recency gate without determining rank [4].
• All candidates support a context length of 256K tokens; the Qwen models use Apache-2.0, while Nemotron-3-Embed-1B-BF16 uses openmdw-1.1 [1][3][4].
• Benchmark status for every candidate: no public benchmark yet (MTEB results, archived date unknown); the ordering does not establish retrieval-quality differences [1][3][4].
• Memory requirements at 4-bit precision are not provided; the listed BF16 weight sizes should not be presented as measured runtime memory or quantized memory requirements [1][3][4].

How do these local embedding models compare on parameters, weight size, context, licenses and benchmark availability?

Why does Qwen's Qwen3-VL-Embedding-8B rank first on the download tiebreak when all candidates lack public benchmarks ?

1. Qwen3-VL-Embedding-8B

Qwen3-VL-Embedding-8B by Qwen ranks first through the popularity tiebreak: with no public benchmark yet for any candidate, its 1,210,187 downloads exceed the other candidates’ totals over the preceding 30 days as of September 25, 2026.[1][3][4] Download activity determines placement here; it does not demonstrate better retrieval quality. Its benchmark status is no public benchmark yet (MTEB results, archived date unknown), so the ordering should not be read as a measured performance ranking.[1][3][4]

Qwen published the model on January 7, 2026, with 8.1B parameters, a 256K-token context window, Apache-2.0 licensing, and 16.3 GB of BF16 weights.[1] For local hardware planning, treat that weight size as the starting point, then allow additional memory for execution and your workload. A device’s advertised memory capacity alone does not establish that the model will fit. The BF16 weight size does not establish a verified 4-bit runtime memory requirement.[1]

Multimodal retrieval is the intended application, described in Qwen’s accompanying paper dated January 8, 2026.[2] Use the model as a candidate for a local retrieval pipeline, then evaluate it against representative queries and documents before committing hardware. The practical caveat is the absence of a public comparative benchmark: popularity supports its position in this shortlist, but cannot establish suitability for your workload.[1][3][4]

2. Qwen3-VL-Embedding-2B

Qwen3-VL-Embedding-2B by Qwen ranks second in this shortlist on the downloads tiebreak, with 1,009,536 downloads over the preceding 30 days as of September 25, 2026 [3]. Qwen’s Qwen3-VL-Embedding-8B has 1,210,187 [1], while nvidia’s Nemotron-3-Embed-1B-BF16 has 519,334 [4]. No public benchmark yet establishes a retrieval-quality ordering among these candidates; popularity determines placement here, not a demonstrated quality advantage.

First published on January 7, 2026, Qwen3-VL-Embedding-2B has 2.1 billion parameters, a context length of 256K tokens, and an Apache-2.0 license [3]. Its BF16 weights occupy 4.3 GB [3], compared with 16.3 GB for Qwen3-VL-Embedding-8B [1]. For local hardware planning, treat the weight size as a starting point, not a complete memory requirement. Allow additional memory beyond the model weights for execution and input processing.

A practical use is evaluating local multimodal retrieval with Qwen’s embedding family when the smaller documented weight footprint suits your machine [2][3][1]. The caveat is unverified comparative retrieval quality: no public benchmark yet is available for this ranking, and the MTEB results archive date is unknown. Evaluate retrieval on your own documents and queries before choosing it for production.

3. Nemotron-3-Embed-1B-BF16

Nemotron-3-Embed-1B-BF16 by nvidia ranks third here under the downloads tiebreak, with 519,334 downloads over the last 30 days as of September 25, 2026 [4], behind Qwen’s Qwen3-VL-Embedding-8B [1] and Qwen3-VL-Embedding-2B [3]. Popularity determines placement because no public benchmark scores these candidates against each other; downloads do not establish retrieval quality. Its July 14, 2026 publication date qualifies it for the release window but does not determine its position [4].

The model has 1.1B parameters, a 256K-token context length, and 2.3 GB of BF16 weights, with an openmdw-1.1 license [4]. For local hardware planning, the published weight size is a starting point, not a complete RAM or VRAM requirement. Allow capacity for runtime allocations alongside the weights. Choose hardware based on a verified memory requirement rather than an assumed quantized footprint.

A practical use is evaluating local embedding workloads where weight storage is a constraint: its published BF16 files are smaller than either ranked Qwen alternative [4][1][3]. The caveat is unverified comparative retrieval quality: no public benchmark yet (MTEB results, archived date unknown). Evaluate retrieval on representative documents and queries before adopting the model; its placement does not establish an accuracy advantage.

What memory information is available, and are quantized VRAM requirements documented?

The available memory figures describe BF16 weight sizes; documented VRAM requirements for 4-bit inference are not provided for any of the candidates [1][3][4].

Qwen’s Qwen3-VL-Embedding-8B has 8.1 billion parameters and BF16 weights totaling 16.3 GB [1]. Qwen’s Qwen3-VL-Embedding-2B has 2.1 billion parameters and BF16 weights totaling 4.3 GB [3]. NVIDIA’s Nemotron-3-Embed-1B-BF16 has 1.1 billion parameters and BF16 weights totaling 2.3 GB [4]. Those figures establish the published weight sizes, but should not be read as measured GPU memory requirements for a running embedding workload.

For a local deployment, use the weight sizes as a starting point for capacity planning, rather than as a guarantee that a model will fit a particular GPU. A comparison should keep published weight size separate from runtime VRAM and label any calculated quantization estimate explicitly as an estimate.

A defensible memory table therefore lists the BF16 weight sizes above and marks 4-bit VRAM requirements as “not documented” for each model [1][3][4]. Before committing hardware, request a runtime measurement for the intended quantized build and workload; avoid treating a conversion from parameter count as a verified deployment requirement.

What does “no public benchmark yet” mean for choosing a local embedding model?

“No public benchmark yet” means there is no public score available to establish a candidate’s retrieval-quality advantage, so treat the ranking as a shortlist for local evaluation. For these candidates, the MTEB status is “no public benchmark yet,” with the results’ archive date unknown. An absent score does not establish poor performance or superior performance.

Qwen’s Qwen3-VL-Embedding-8B appears first because download popularity breaks the benchmark tie, without establishing a measured retrieval-quality advantage [1][3][4]. The remaining order is Qwen’s Qwen3-VL-Embedding-2B, then NVIDIA’s Nemotron-3-Embed-1B-BF16 [3][4]. Their respective download counts over the preceding 30 days were 1,210,187, 1,009,536 and 519,334 as of September 25, 2026 [1][3][4]. Release recency determines eligibility; neither a newer release nor a smaller weight footprint changes that ordering.

Hardware suitability still matters when choosing what to evaluate. Published BF16 weight sizes are 16.3 GB, 4.3 GB and 2.3 GB, respectively [1][3][4]. Those figures describe weight files; they do not establish total runtime memory or quantized memory requirements. Avoid treating them as guarantees that a model will fit your machine.

Evaluate each candidate against the same documents, queries and relevance judgments. Measure retrieval quality, indexing time, query latency and peak memory on your hardware. Choose according to your workload’s requirements, and label any model-card benchmark “self-reported” rather than presenting it as an independent comparison.

Which licenses apply to these local embedding models?

Qwen’s Qwen3-VL-Embedding-8B (Qwen) and Qwen3-VL-Embedding-2B (Qwen) list Apache-2.0 as their license [1][3]; nvidia’s Nemotron-3-Embed-1B-BF16 (nvidia) lists openmdw-1.1 [4]. Keep that distinction explicit when choosing a model for a local deployment: the candidates do not all carry the same license label [1][3][4].

Both Qwen model cards identify Apache-2.0 [1][3]. Switching between those Qwen variants therefore preserves the listed license, although the repository recorded in your deployment documentation should match the weights you actually use. Record each model separately rather than documenting only the Qwen family name.

The nvidia model’s listed license is openmdw-1.1 [4]. Treat that exact identifier as a separate item in your deployment review. Do not copy the Qwen license entry into a shared configuration or inventory when adding the nvidia model.

For an engineering handoff, include the selected repository, its listed license and a link to the corresponding model card. Before modifying weights, redistributing them or packaging a model into a product, check the applicable license text for your intended use. Keep the license decision separate from retrieval evaluation and hardware sizing.

Frequently Asked Questions

Which local embedding model ranks first in this comparison?

Qwen’s Qwen3-VL-Embedding-8B ranks first because all candidates lack public benchmark results and its 1,210,187 downloads over the last 30 days, measured on September 25, 2026, place it first under the popularity tiebreak [1][3][4]. Popularity determines the ordering here; it does not establish superior retrieval quality. Treat this placement as a shortlist position rather than a demonstrated performance advantage.

What is the ranking order, and why?

The order is Qwen’s Qwen3-VL-Embedding-8B, Qwen’s Qwen3-VL-Embedding-2B, then nvidia’s Nemotron-3-Embed-1B-BF16 [1][3][4]. Their respective downloads over the last 30 days were 1,210,187, 1,009,536, and 519,334 as of September 25, 2026 [1][3][4]. All pass the release-window gate, and no public benchmark separates them. Download counts therefore break the tie; release recency and weight size do not determine placement.

How much memory do I need to run these models locally?

Published BF16 weight sizes are 16.3 GB for Qwen3-VL-Embedding-8B [1], 4.3 GB for Qwen3-VL-Embedding-2B [3], and 2.3 GB for Nemotron-3-Embed-1B-BF16 [4]. Parameter counts are 8.1B [1], 2.1B [3], and 1.1B [4], respectively. A verified quantized runtime memory requirement is unavailable. Use the weight sizes as storage information, rather than a guarantee that deployment will fit within an equivalent amount of RAM or VRAM.

Which of these models supports long documents?

Qwen3-VL-Embedding-8B, Qwen3-VL-Embedding-2B, and Nemotron-3-Embed-1B-BF16 each list a context length of 256K tokens [1][3][4]. Context length therefore does not distinguish these candidates. The listed limit also does not establish a retrieval-quality ranking for long documents. Evaluate document handling separately from the popularity ordering, and avoid treating a context specification as a benchmark result.

What licenses do these embedding models use?

Qwen3-VL-Embedding-8B and Qwen3-VL-Embedding-2B use Apache-2.0 [1][3]. Nemotron-3-Embed-1B-BF16 uses openmdw-1.1 [4]. License choice is a deployment consideration separate from ranking. Check the applicable terms against your intended use, especially before redistributing weights or including them in a product. A shared place on an open-weight shortlist does not mean the models have identical license terms.

Do public benchmarks show which model has better retrieval quality?

The benchmark status for Qwen3-VL-Embedding-8B, Qwen3-VL-Embedding-2B, and Nemotron-3-Embed-1B-BF16 is “no public benchmark yet” for MTEB results, with archived date unknown [1][3][4]. No dated public score establishes a quality winner among these candidates. Any model-card benchmark must be labeled self-reported. The popularity ordering should therefore guide shortlist presentation without being presented as measured retrieval performance.

Sources

1. Qwen/Qwen3-VL-Embedding-8B model card (Hugging Face) — 2026-09-25
2. Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking — 2026-01-08
3. Qwen/Qwen3-VL-Embedding-2B model card (Hugging Face) — 2026-09-25
4. nvidia/Nemotron-3-Embed-1B-BF16 model card (Hugging Face) — 2026-09-25
5. Compact Language Models via Pruning and Knowledge Distillation — 2024-07-19