NVIDIA

A streaming version of Nemotron ASR, the successor of the beloved Parakeet series.

The big version of the Nemotron series, which uses LatentMoE to be even faster than comparable models. Just like the other Nemotron models, the vast majority of the data is open source. And, to top it all off: NVIDIA commits to using the OpenMDW license, which is tailored specifically for model weights (and data) and drops its custom license. While MIT and Apache are in the same spirit as OpenMDW, only the latter really covers model weights, while the former are software licenses that do not really apply to model weights.

An omni version of Nemotron Nano, supporting video, audio, image, and text as inputs. As usual for NVIDIA, they describe and release most of the data and the generation methods used to obtain it.

An experimental model which can be used in three different modes: autoregressive, diffusion, and self-speculation. Each of these modes is suitable for a different use case.

Small OCR models by NVIDIA, coming in both English-only and multilingual variants.

A pruned expert version of GPT OSS 120B. It also replaces some global attention layers with window attention. Puzzle is "a post-training neural architecture search (NAS) framework, with the goal of significantly improving inference efficiency for reasoning-heavy workloads while maintaining or improving accuracy across reasoning budgets."

A compressed version of NVIDIA-Nemotron-Nano-9B-v2, which itself is a compressed version of NVIDIA-Nemotron-Nano-12B-v2. Nvidia has been pushing this direction more than anyone else with open models.

A streaming, English-only version of Nemotron, one of the best speech-to-text models.

The long-awaited mid-sized model from NVIDIA is finally here: 120B total params with 12B active, a 1M context window, and support for multiple popular languages. Furthermore, the model is based on LatentMoE and uses NVFP4 during pre-training, which is a first for open models. Like other things from NVIDIA, it comes with an in-depth tech report plus pre-training and post-training datasets, with the vast majority of the data being openly released.

A speech generation model with support for 5 voices and 7 languages per voice.

A multilingual, multimodal and multi-vector embedding model.

A multimodal, single vector embedding model combining Llama 3.2 1B and SigLip 2.

An action model for video games.

As luck would have it, NVIDIA released an update to their Nemotron series right after our year in review tierlist for 2025. Similar to other NVIDIA models, the vast majority of the data is released openly. Furthermore, they continue with the Mamba2-Transformer architecture, but make it a MoE as well. And to top it all off: They also announce two more sizes, slated for a release in H1 2026 (likely on the earlier side): Super, ~100B-A10B and Ultra, ~500B-A50B, which will use Latent MoE and multi-token prediction (MTP). 2026 will be an exciting year!

A small detection model for personal identifiable information.

NVIDIA is one of the very few Western companies which continues to release a lot of open models and is often seen as a guest in the Artifact series. It is almost ironic that a company, which is arguably one of the most GPU-rich companies out there, continues to release models which run on limited hardware. Nemotron "Nano", which comes in 9B and 12B sizes, is a hybrid model consisting of both attention and mamba layers. It is also hybrid in the sense that it supports a reasoning and non-reasoning model. They also share (the majority of the) the pre-training data. It is hard to say "don't sleep on NVIDIA" as it is arguably one of the most important companies during the AI boom, but them continuing to release capable, small models AND open data with detailed papers is really commendable. On top of all of that, they make the model available for free on OpenRouter.

A bigger version of NVIDIA's FastConformer architecture, which also powers parakeet. The architecture of Canary differs slightly. The bigger size also helps with more accurate transcriptions.

There are a few models which completely redefine their space and push the boundaries. For open models, this is even harder as the competition from closed models is fierce. However, Parakeet fits these criteria and will probably be overlooked by a lot of people. But for me (Florian), it has redefined how I use my MacBook. Last episode, I wrote the following about Qwen3 4B Instruct: I have started using it locally for simple tasks (like translation), as the model is really capable and the overall latency is faster than sending requests to the cloud. The Qwen model gets complemented by this model perfectly when used in apps such as MacWhisper: Parakeet is used for the transcription, while Qwen 4B is used to clean up the raw transcripts. Both models are blazingly fast (and therefore beat cloud-based models in terms of overall latency), yet accurate. This is the big selling point of local models on device: they are tiny niche models that are as capable as more general, closed models. And Parakeet fits this description perfectly. Since it launched, I am shifting more towards using my voice to write emails, long prompts for ChatGPT or Codex or this very blog post. For me as a multilingual speaker, I felt left behind with the previous version, which was English only.

Nvidia is not just one of the few companies that still use Llama as a base model, it is also one of the few American model providers with a cadence and quality that matches Chinese companies, as this release proves yet again. Interestingly, it is way less censored than its predecessor as evaluated by SpeechMap:

An audio model combining the speed of NVIDIA's Canary models with the quality of Qwen.

A new version of NVIDIA's reasoning model, using the same prompts as the previous version, but with traces generated by R1 0528.

A multimodal, multi-vector retrieval model using SigLIP and Llama 3.2. However, it is released under a noncommercial license.

NVIDIA is one of the few companies still building upon Llama in their Nemotron series. This model combines Llama3.1 8B with their own vision encoder.

Another reasoning model by NVIDIA, building upon R1-Distill-Qwen-1.5B.

A robotic reasoning model from NVIDIA.

Building upon R1-Qwen 14B, AceReason uses two RL phases: One exclusively for math, followed by one for code. Like others they use GRPO.

A pruned version of Llama 3.1 8B.

Multimodal model for robotic tasks.

English speech transcription model using the FastConformer architecture, released under CC-BY-4.0.

A reasoning model based on distilled R1 traces.

A math model, based on Qwen 32B. The paper goes into more detail for the data creation, which needed a lot of work to get another model (LIMO-Qwen 32B) to output proper tool calls. They also used a lot of different reasoning models to create the data.

A model to describe parts or points of images and videos. For this purpose, it uses both a crop of the area of interest, as well as the whole image as context.

A hybrid transformer-mamba model trained on a whopping 20T tokens. The report goes into greater detail. The models perform on the same level as an attention-only variant, while being more than twice as fast. The long-context benchmarks also look competitive, making this architecture a serious contender to replace models which use sliding window attention or similar variants. The released models, however, support only 8K context.

A single encoder image encoder model trained for a lot of tasks, released under a permissive license.

The first hybrid Transformer-Mamba vision model. Similar to text-only models, they find that the combination of attention and mamba layers is superior compared to only using Mamba or attention layers. The accompanying paper goes into more detail, including ablation studies.

A foundation model for humanoid robots. The overall design is similar to Helix from Figure: A VLM captures the environment and the prompt to plan the task. A second model, in this case a diffusion transformer, uses this information to output the actions to be executed by the robot.

A speech-to-text model for four languages: English, German, French, Spanish. It performs better than Whisper large-v3, while being smaller and substantially faster due to its architecture using only few transformer layers.

NVIDA has also released a new version of their EAGLE VLMs by using Qwen2.5 as the LLM backend with SigLIP as the vision encoder. The models are available under a non-commercial license.

NVIDIA has released "World Foundation Models", which are models capable of outputting videos based on various input modalities (text, images, frames). The models are trained by focusing on physically accurate simulations to enable robots. For the evaluation, they also tested their models against a physics engine.

A model Nvidia recently fine-tuned that scored extremely high on LLM-as-a-judge evals like MT Bench and ArenaHard. The final jury will be LMSYS, but it's exciting to see fine tunes continuing to climb upwards. This was trained on the HelpSteer2 data. At the same time, we need actually good models and not just high evaluations.

Two models Nvidia has trained by distillation and pruning from larger models. The numbers are good for these models, but I'm yet to see big adoption. I'm planning to write an entire post on this eventually. It feels important, and early.

Nvidia has entered the VLM game! Performance isn't crazy good, but good to have more options.

Probably the strongest open-weights reward model I know of right now for multiple uses.

Two models Nvidia has trained by distillation and pruning from larger models. The numbers are good for these models, but I'm yet to see big adoption. I'm planning to write an entire post on this eventually. It feels important, and early.

Nvidia's base model that is pruned from a larger model - i.e. they manage to remove layers and still have the performance be effective.

A general question answering and RAG model from Nvidia. People tend to sleep on their fine-tuning teams, this looks solid (figure below)!