NVIDIA releases Nemotron 3.5 Rerank and Embed, alongside CDF Sim2Real SMT Solderlight

What was released and when

NVIDIA published four artificial intelligence models on October 20, 2026. These include two embedding variants (Nemotron 3.5 Embed 8B BF16 and Nemotron 3.5 Embed 1B BF16), a reranking model (Nemotron 3.5 Rerank 8B BF16), and an image editing model specific to the electronics industry (CDF Sim2Real SMT Solderlight).

The release was made simultaneously in official repositories, confirming the immediate availability of weights in BF16 format. No deferred launch date or beta phase was announced; the models are ready for download and deployment by end users.

The simultaneous availability of weights in BF16 format across official repositories underscores a direct-to-user distribution strategy, bypassing intermediate commercial gateways. This immediate accessibility allows end users to initiate deployment workflows without waiting for phased rollouts or beta testing periods, although the absence of a managed API means that the operational burden of infrastructure setup remains entirely with the deploying organization.

The inclusion of both an 8B and a 1B parameter variant in the embedding family suggests a deliberate design choice to accommodate diverse hardware constraints. While the larger model offers higher dimensional output, the smaller variant provides a pathway for integration in environments where computational resources are limited, ensuring that the multimodal capabilities of the Nemotron 3.5 family are not restricted solely to high-end GPU clusters.

Variants, modalities, and access

The Nemotron 3.5 family models are multimodal, accepting both text and page images as input. The reranking model scores the relevance of each query-document pair, while the embedding models generate normalized vectors (4,096 dimensions for the 8B and 2,048 for the 1B).

Access is direct through the official repositories. There is no NVIDIA-managed commercial API for these specific models, which means the user must manage their own compute infrastructure. The OpenMDW 1.1 license allows commercial use, subject to its terms and the preservation of copyright notices.

The multimodal nature of these models, accepting both text and page images, represents a significant architectural shift for retrieval systems. By processing mixed content in a single stage, the models eliminate the need for complex preprocessing pipelines that traditionally separated textual and visual data streams. This integration simplifies the development of search interfaces that can handle technical documentation containing both descriptive text and schematic diagrams.

The licensing structure, which combines the OpenMDW 1.1 license for the new models with the retention of Apache 2.0 conditions for the underlying Ministral bases, creates a layered legal framework. Users must carefully navigate these terms to ensure compliance, particularly regarding attribution and the specific restrictions or permissions associated with commercial use, as the interaction between the two licenses may impose additional obligations beyond standard open-source practices.

What changes for the user

For software engineering teams, the main novelty is the ability to process mixed documents (text and image) in a single retrieval or reranking stage. This simplifies search architectures that previously required separate pipelines for text and vision.

The CDF Sim2Real SMT Solderlight model is more specific: it is designed to transform synthetic crops of printed circuit board (PCB) components into an optical inspection lighting style. It is not a quality decider, but a visual preprocessing tool that requires a subsequent inspection system and human review.

For software engineering teams, the consolidation of text and image processing into a unified retrieval or reranking stage reduces the complexity of search architectures. Previously, systems required separate pipelines for handling textual queries and visual content, often leading to inconsistencies in relevance scoring. The new models allow for a more coherent evaluation of document relevance, where the context provided by diagrams can directly influence the ranking of associated textual segments.

The CDF Sim2Real SMT Solderlight model serves a distinct role within this release, focusing on the electronics industry's specific need for visual preprocessing. By transforming synthetic PCB component crops into an optical inspection lighting style, it acts as a bridge between simulation and real-world inspection. However, its function is strictly preparatory; it does not replace the need for downstream quality determination systems or human oversight, emphasizing its role as a tool for enhancing visual fidelity rather than making autonomous decisions.

Pricing, limits, and what remains to be verified

There are no published API fees. The cost to the user is limited to their own infrastructure: GPU, memory, storage, and operation. The 8B parameter models require more powerful hardware than the 1B variant, which is designed for resource-constrained environments.

An important limit is the context of 32,768 tokens shared between text and image. Additionally, images are processed with a default longest side of 1,120 pixels, expandable to 1,540. No independent numerical benchmarks have been published for these specific checkpoints, so performance evaluation must be performed by the user in their own context.

The cost structure for these models is defined by the user's own infrastructure requirements rather than by service fees. Since no commercial API is provided, the financial implications are tied directly to the hardware necessary to run the models, including GPU power, memory capacity, and storage. The 8B parameter models demand more robust hardware, while the 1B variant is tailored for environments with tighter resource constraints, allowing organizations to scale their deployment based on their specific computational budget.

Technical limitations, such as the shared context window of 32,768 tokens and the default image processing resolution of 1,120 pixels, impose boundaries on the volume and detail of data that can be processed in a single inference. The lack of independent numerical benchmarks for these specific checkpoints means that performance validation is a critical step for users, who must conduct their own evaluations to ensure that the models meet their specific latency and accuracy requirements in their operational context.

Technical considerations and deployment

Deployment requires attention to software dependencies. The official examples use Transformers version 5.20.0 with Flash Attention 2, while vLLM 0.32.0 requires an environment with a lower version of Transformers. It is crucial to review any remote code (trust_remote_code) before executing it in production.

For the CDF Sim2Real SMT Solderlight image model, training was performed with an area of 262,144 pixels (512x512). Using very different resolutions may degrade the quality of the output. The model derives from Qwen-Image-Edit with a fused LoRA adjustment, meaning it is delivered as a complete pipeline, not as an independent adapter.

Deployment of these models requires careful management of software dependencies, particularly regarding the compatibility between Transformers and vLLM versions. The official examples utilize specific versions of these libraries, and deviations from these configurations may lead to operational issues. Additionally, the use of remote code execution features necessitates a thorough security review before production deployment to mitigate potential risks associated with unverified code paths.

The CDF Sim2Real SMT Solderlight model, derived from Qwen-Image-Edit with a fused LoRA adjustment, is delivered as a complete pipeline rather than an independent adapter. This design choice simplifies integration but requires users to adhere to the training resolution of 512x512 pixels to maintain output quality. Significant deviations from this resolution may result in degraded performance, highlighting the importance of aligning input data specifications with the model's training parameters.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.