LoginSign Up
← Back to all sparks
N

NVIDIA NeMo

AI-ASSISTANTS
Velocity3.8

Framework for building and training generative AI and speech models

developer.nvidia.com/nemo ↗

NeMo split itself apart: the flagship repo is now a speech toolkit and nothing else.

speech-aiasrttsrepo-splittechnical-debtbreaking-changes
◆Current state
NeMo has spent the last six months on a controlled demolition. The 2.7.0 notes warned that avlm, diffusion, llm, multimodal, nlp, speechlm, vision and vlm collections would be removed; NeMo Speech 3.0 executed it, splitting the repository, renaming it to NVIDIA-NeMo/Speech and moving everything non-speech to sibling repos. The release removed 800k lines of deprecated code, moved to uv for installs, cut dependencies and shipped lighter containers. Patch releases in between were security fixes and CUDA binding repairs.
◆Where it's heading
This is a scope decision, not a cleanup. NeMo is trading its position as a general-purpose model framework for a defensible one as the speech toolkit — ASR, TTS, speaker tasks and SpeechLM — and accepting a hard migration for everyone else. The feature work that did ship in 2.7.0 points the same way: streaming speech translation, per-stream phrase boosting, and new streaming ASR and multilingual TTS models.
◆Prediction
With the split done, expect the next releases to be speech-model drops rather than framework changes, and the separated repos to start versioning independently.

◆Recent moves

  1. 2d ago

    NVIDIA NeMo Speech 3.0

    ⚡ SPARK

    The narrowing announced in 2.7.0 lands in full. NeMo is now a speech repository, and the 800k lines removed plus the uv migration are the cost of making that focus real rather than nominal.

    View source ↗
  2. 3mo ago

    Security patch release with restricted unpickling

    A security-driven patch on the 2.7 line, notably switching to a RestrictedUnpickler in the checkpoint loading path. Housekeeping ahead of the split rather than new capability.

    View source ↗
  3. 4mo ago

    Patch: numba-cuda and cuda-python installation fixes

    A two-commit patch repairing CUDA dependency installation and usage. No user-facing behaviour change beyond installs that now work.

    View source ↗
  4. 4mo ago

    Patch: CUDA graphs binding fix

    A single cherry-picked fix for cuda-python usage in CUDA graphs. The smallest possible release.

    View source ↗
  5. 5mo ago

    Streaming speech translation, new models, and the split announced

    Substantive speech work — streaming speech translation, per-stream phrase boosting in transducer decoding, and the nemotron-speech-streaming-en-0.6b and magpie_tts_multilingual_357m model releases — shipped alongside the notice that nine non-speech collections would be removed. In hindsight this release is where the 3.0 direction was set.

    View source ↗
  6. 6mo ago

    Security patch and torch weights-only load hardening

    A patch on the 2.6 line addressing known security issues, pinning torch to 2.6 or later and removing weights_only=False overrides. Part of the same checkpoint-loading hardening that continued into 2.7.3.

    View source ↗