NVIDIA Nemotron Timeline: Model Release Dates & History

NVIDIA released the first Nemotron-3 8B language models in 2023. By 2026, the family included the 550B-parameter Nemotron 3 Ultra and specialist releases for speech, vision, and document understanding. NVIDIA also developed Llama-based reasoning models and the hybrid Mamba-Transformer architecture used in the Nemotron 3 generation.

As the Nemotron family grew, release dates became harder to track from model names alone. The original Nemotron-3 and the newer Nemotron 3 belong to different generations, and NVIDIA sometimes announced a family months before releasing its individual models. The timeline records the major announcements and public launches through October 2026, with the developments introduced at each milestone.

Last updated: October 26, 2026.

2026: Nemotron 3, Nemotron 3.5, and Specialist Releases

Date	Milestone
Oct. 23, 2026	Nemotron 3.5 Bolt E4B 4GB is released for compact conversational agents. It uses Gemma 4 E4B with Nemotron post-training and includes conversation, tool-use, and configurable reasoning controls.
Oct. 22, 2026	NemotronLabs RE-USE is released for multilingual speech enhancement. The 9.6-million-parameter model restores degraded speech, with a separate checkpoint for preserving laughter.
Oct. 20, 2026	Nemotron 3.5 Embed and Rerank are released for multilingual and multimodal retrieval. The group includes 1B and 8B embedding models and an 8B reranker for text and document-page images.
Oct. 17, 2026	Nemotron 3.5 Parse is released for document extraction. The 631M-parameter parser extracts text, layout, reading order, tables, and charts from document images.
Oct. 16, 2026	Nemotron Labs Dory1.0 4B is released as a recurrent language model. Repeated use of shared layers lets the same checkpoint use different amounts of inference computation.
Oct. 15, 2026	Nemotron 3.5 Super VL launches with 120B total parameters and 12B active per token. Built on Nemotron 3 Super with distillation from Nemotron 3 Ultra, it handles text, images, documents, and video alongside reasoning and tool use.
Oct. 15, 2026	Nemotron 3.5 Bolt 74B-A9B 24GB is released for local agent workloads. Compression and quantization adapt the Super VL model to a 24GB deployment target.
Oct. 14, 2026	Nemotron 3.5 Super and Lightning are described in an NVIDIA Research report on hybrid Mamba-Transformer models, long-context reasoning, and speculative decoding. The Super VL weights launch the following day.
Sep. 23, 2026	Nemotron 3 Diarization is released for identifying speaker turns in live and recorded audio. It supports up to eight speakers and extends the earlier four-speaker Sortformer approach.
Aug. 11, 2026	Nemotron 3.5 Lightning launches with 30B total parameters and 3B active per token. Distilled from Nemotron 3 Super, it targets repeated tool execution and specialized tasks in long-running agents.
Aug. 3, 2026	Nemotron Parse 2.0 is released for structured document extraction. It continues the specialist parser line with text, table, layout, and reading-order outputs.
Aug. 3, 2026	NemotronLabs VoiceChat 11B is released as a full-duplex conversational speech model. It combines a Nano v2 language backbone with speech understanding, generation, and a tool-calling output channel.
Jul. 16, 2026	Nemotron 3 Embed models are released in 1B and 8B sizes. The text-retrieval line supports search across enterprise documents, code, and agent memory.
Jun. 4, 2026	Nemotron 3 Ultra launches with 550B total parameters and 55B active per token. It expands the generation’s largest tier for complex reasoning and agent orchestration.
Jun. 4, 2026	Nemotron 3.5 ASR Streaming 0.6B is released for multilingual streaming speech recognition. It extends the earlier English streaming-ASR line.
Jun. 2, 2026	Nemotron 3.5 Content Safety is released. It combines the text-and-image moderation of Nemotron 3 Content Safety with custom-policy reasoning from the earlier Content Safety Reasoning model.
May 19, 2026	Nemotron-Labs-Diffusion releases base, instruction, and vision-language models. The research family supports autoregressive decoding, parallel diffusion decoding, and self-speculation within the same model.
Apr. 28, 2026	Nemotron 3 Nano Omni launches with text, image, audio, and video understanding. The 30B-A3B model supplies multimodal perception and reasoning for agent workflows.
Apr. 15, 2026	Nemotron OCR v2 becomes available through NVIDIA Build and NGC. English and multilingual variants combine text detection, recognition, and document-layout relationships.
Mar. 19, 2026	Nemotron-Cascade-2-30B-A3B receives its Hugging Face weights release. The downloadable checkpoint follows the March 16 research introduction.
Mar. 16, 2026	Nemotron 3 Nano 4B is released for compact conversational and tool-using applications. It is a smaller hybrid Mamba-Transformer checkpoint, distinct from the 30B-A3B Nano model.
Mar. 16, 2026	Nemotron 3 Content Safety is released for multilingual text and image moderation. Built on Gemma 3 4B, it adds a multimodal model to the safety branch.
Mar. 16, 2026	NVIDIA announces Nemotron 3 Omni and VoiceChat at GTC. Omni targets combined audio, visual, and language understanding. VoiceChat targets real-time conversations with simultaneous listening and speaking.
Mar. 16, 2026	NVIDIA announces the Nemotron Coalition. Its planned open model will underpin the future Nemotron 4 family. The planned family is separate from the Nemotron-4 models released in 2024.
Mar. 16, 2026	Nemotron-Cascade 2 is introduced as a 30B mixture-of-experts reasoning model with 3B active parameters. It builds on the Nemotron 3 Nano base model and adds multi-domain on-policy distillation.
Mar. 11, 2026	Nemotron 3 Super launches with 120B total parameters and 12B active per token. It adds LatentMoE, multi-token prediction, and a one-million-token context window to the larger Nemotron 3 tier.
Feb. 24, 2026	Nemotron-Terminal is introduced in a technical report on training models for terminal tasks. The family includes Qwen3-based 8B, 14B, and 32B models for autonomous command-line interaction.
Feb. 17, 2026	Nemotron Parse v1.2 is released with an updated document-prompt format, text extraction from embedded images, and improved reading order for tables, figures, and captions.
Feb. 17, 2026	Nemotron Nano 9B v2 Japanese is released. Additional Japanese training adapts the Nano 2 model for reasoning, instruction following, and tool calling in Japanese.
Jan. 26, 2026	Nemotron ColEmbed V2 models are released for visual document retrieval. Their late-interaction design compares multiple query and document representations. This extends the retrieval branch beyond single-vector embeddings.
Jan. 5, 2026	Nemotron Speech Streaming English 0.6B is released for streaming transcription. The model adds low-latency automatic speech recognition to the Nemotron speech line.

2025: Llama Nemotron, Hybrid Reasoning, and Nemotron 3

Date	Milestone
Dec. 18, 2025	Llama Nemotron Embed VL and Rerank VL 1B v2 are released for visual retrieval. They add document-page images and combined image-text inputs to embedding and relevance-ranking workflows.
Dec. 15, 2025	Nemotron 3 is announced, and Nano 30B-A3B becomes available with open weights. The hybrid Mamba-Transformer mixture-of-experts model starts the new generation. Super and Ultra are announced for later release.
Dec. 15, 2025	Nemotron-Cascade is introduced with general-purpose reasoning models trained with cascaded reinforcement learning. The branch includes 8B and 14B checkpoints with thinking-only and unified reasoning variants.
Nov. 24, 2025	Nemotron-Flash is introduced in a research paper as a 1B and 3B hybrid language-model family. Its architecture search targets decoding latency, with base and instruction-tuned models.
Nov. 20, 2025	Nemotron Elastic is introduced in a research report. Its reasoning models contain nested submodels that share weights and can be extracted for different deployment budgets without additional training.
Nov. 17, 2025	Nemotron Parse 1.1 receives its Hugging Face weights release after its October announcement. The parser produces text, layout classes, bounding boxes, and reading order from document images.
Oct. 28, 2025	Nemotron Nano 2 VL launches as a 12B vision-language model alongside Llama 3.1 Nemotron Safety Guard 8B v3. NVIDIA announces Nemotron Parse 1.1, expands its retrieval models, and previews the future Nemotron Nano 3 model.
Oct. 23, 2025	Nemotron OCR v1 is released on Hugging Face alongside page-element, graphic-element, and table-structure models. The group adds specialist extraction components for document-processing pipelines.
Oct. 21, 2025	Llama-Embed-Nemotron-8B is released for text retrieval. It adds a larger embedding model to the Nemotron retrieval lineup.
Oct. 1, 2025	Omni-Embed-Nemotron 3B is released for cross-modal retrieval. It embeds text, images, audio, and video into a shared representation for search.
Aug. 29, 2025	Nemotron Nano 12B v2 is released as a post-trained model. It complements the earlier 9B model and 12B base checkpoint in the Nano 2 branch.
Aug. 18, 2025	Nemotron Nano 2 launches with the 9B v2 reasoning model and 12B base checkpoint. Its hybrid Mamba-Transformer design follows Nemotron-H and supports controllable reasoning.
Jul. 25, 2025	Llama Nemotron Super 49B v1.5 is released with additional post-training for reasoning, coding, instruction following, and tool use. It updates the earlier Super 49B model.
Jul. 16, 2025	OpenReasoning-Nemotron checkpoints are released in 1.5B, 7B, 14B, and 32B sizes for mathematical, coding, and scientific reasoning. They are post-trained derivatives of Qwen models.
Jun. 27, 2025	Nemotron releases a reward-model suite based on Llama 3.3 and Qwen. It includes response-scoring models and a Super 49B generative reward model for evaluating candidate answers.
Jun. 16, 2025	AceReason-Nemotron 1.1 7B is released. The update combines supervised fine-tuning with reinforcement learning for math and code reasoning.
Jun. 11, 2025	Mistral-Nemotron is available through NVIDIA NIM for coding, instruction following, and tool calling. The Mistral-based model extends the Nemotron lineup beyond its Llama-based branch.
Jun. 6, 2025	Nemotron-H adds 8B and 47B reasoning checkpoints with 128K context. These models support both reasoning and direct-response modes on the hybrid Mamba-Transformer architecture.
Jun. 3, 2025	Llama Nemotron Nano VL 8B V1 launches for document understanding. The vision-language branch adds image, chart, table, and text interpretation to the Nano lineup.
May 22, 2025	AceReason-Nemotron is introduced in a technical report on math and code reasoning. The research applies reinforcement learning to DeepSeek-R1-distilled Qwen models at 7B and 14B.
May 20, 2025	Llama Nemotron Nano 4B v1.1 is released as a smaller reasoning model. It is derived from a compressed Llama 3.1 8B model and supports chat, retrieval, and tool calling.
Apr. 25, 2025	OpenCodeReasoning-Nemotron models are released in 7B, 14B, and 32B sizes. The Qwen-based checkpoints extend Nemotron’s specialist code-generation branch.
Apr. 23, 2025	OpenMath-Nemotron models are released in 1.5B, 7B, 14B, and 32B sizes. They adapt Qwen-based models for mathematical reasoning with the OpenMathReasoning training data.
Apr. 14, 2025	Nemotron-UltraLong-8B is introduced with one-, two-, and four-million-token context variants. The models are based on Llama 3.1 8B Instruct.
Apr. 7, 2025	Llama Nemotron Ultra 253B v1 is released. Derived from Llama 3.1 405B, it adds the largest model in the initial Llama Nemotron reasoning lineup.
Mar. 21, 2025	Nemotron-H is introduced as a hybrid Mamba-Transformer family. It establishes the hybrid architecture line that later supports Nemotron Nano 2.
Mar. 18, 2025	Llama Nemotron Nano 8B and Super 49B are released with controllable reasoning. Nano builds on Llama 3.1, while Super uses Llama 3.3 with architecture search and distillation.
Jan. 6, 2025	NVIDIA announces the Llama Nemotron family, with Nano, Super, and Ultra tiers for agent workloads. Individual model releases follow later in 2025.

2024: Nemotron-4 and Llama-Based Models

Date	Milestone
Oct. 23, 2024	Nemotron-4-Mini-Hindi-4B becomes available as an NVIDIA NIM microservice. The Hindi-English model is derived from Nemotron-4 15B through compression and additional training.
Oct. 15, 2024	Llama-3.1-Nemotron-70B-Instruct becomes publicly available as an NVIDIA post-trained version of Llama 3.1 70B. It extends the line’s instruction-following and conversational models.
Sep. 23, 2024	Llama-3.1-Nemotron-51B-Instruct launches as a compressed derivative of Llama 3.1 70B. Neural architecture search and block distillation reduce the model’s inference requirements.
Aug. 20, 2024	Nemotron-Mini-4B-Instruct is announced for on-device conversation, roleplay, retrieval, and function calling. Its compressed 4B design brings the Nemotron line to local digital-character applications.
Jun. 14, 2024	Nemotron-4 340B launches with Base, Instruct, and Reward models. The release establishes a model pipeline for generating synthetic training data and evaluating response quality.
Feb. 26, 2024	Nemotron-4 15B is described in its first public technical report. The dense multilingual model establishes the earlier Nemotron-4 research line, separate from the later 340B release.

2023: The Original Nemotron-3 8B Family

Date	Milestone
Nov. 15, 2023	Nemotron-3 8B is introduced for enterprise language applications. The original family includes base, supervised chat, RLHF chat, SteerLM chat, and question-answering variants.

FAQs

Is NVIDIA Nemotron the same as NVIDIA NeMo?

Nemotron is a model family. NeMo provides software for developing, training, evaluating, and deploying AI models.

Do Nano, Super, and Ultra refer to fixed model sizes?

No. Their parameter counts and architectures change across lineups. A Nano name alone does not identify the model’s generation, size, or architecture. Check the complete model name, including its generation, parameter count, and version suffix.

When did NVIDIA release Nemotron 3?

NVIDIA announced Nemotron 3 on December 15, 2025, and released the Nano 30B-A3B model that day. Super followed on March 11, 2026, and Ultra on June 4, 2026. These belong to a later generation than the original Nemotron-3 8B family from November 2023.