NVIDIA-NemotronLabs-VoiceChat-11B-Spark

NVIDIA NemotronLabs VoiceChat 11B — DGX Spark community quantization

This is a community conversion of nvidia/NVIDIA-NemotronLabs-VoiceChat-11B at immutable parent revision fb0f94eaf4d03ddc430f39565229393fa1b50c26. It is not an NVIDIA release. It targets one NVIDIA DGX Spark (GB10, Linux ARM64) and the runtime in github.com/pipecat-ai/nemotron-voicechat-dgx-spark.

Contents

• nano/: calibrated symmetric GPTQ W8A16, group 128, for 105 MLP, Mamba, and text-head tensors. The function head, embeddings, and attention projections remain full precision. K=15,680 down projections are zero-padded to 15,744 for group-128 Marlin execution.
• eartts/: symmetric per-output-channel W8 weights with FP32 activations and accumulation. The qualified runtime configuration uses a 1,500-position sliding window.
• calibration/replay.tar: exact fused-embedding calibration/evaluation replays in a deterministic archive; calibration/manifest.json inventories and hashes every archived file. The corpus contains no raw user recordings.
• manifests/release.json: SHA-256 and byte count for every release file.
• reports/: portable conversion evidence.

The ordinary setup path downloads this release. The source repository also contains the deterministic production conversion code, corpus selection, 100 GiB disk guard, dual clean-process conversion check, component gates, and exact reproduction command.

Qualification

Production Candidate 1 passed the held-out Nano control-token gate, EarTTS decode with independent Parakeet transcription, real browser/Pipecat SmallWebRTC multi-turn tests, and function calling. Its two-frame mean passed the 80 ms cadence target and its two-frame p95 was a documented near miss near 82 ms. Browser playback uses a small adaptive jitter buffer because delivery jitter can exceed one 80 ms audio slice.

The current public runtime passed paced 12,000-frame qualification on both the downloaded artifact and a byte-identical public-source reproduction. Late-block queue slopes were 0.289 and 0.375 ms/minute, with peak implied playback debt of 382 and 400 ms respectively; the earlier linear EarTTS drift did not recur. All 18 typed turns in each successful sustained run produced non-silent audio that passed independent ASR. Sessions close gracefully at the declared model frame limit.

The qualified platform identity is:

• NVIDIA DGX Spark / GB10, compute capability 12.1
• NVIDIA driver 580.142
• Linux kernel 6.17.0-1014-nvidia
• NVIDIA PyTorch 25.12 lineage

Other hardware and software combinations are not qualified by this release.

Known limitations

The upstream model can self-talk, repeat canned replies, choose or skip the wrong tool, mis-speak tool results, drop transcript words, or occasionally produce a silent response. The deployment adds bounded watchdog and recovery behavior but cannot eliminate model-level errors. One retained source-runtime campaign also encountered a malformed late function-call recovery that disabled two-frame drafting and failed the queue-debt gate; a fresh-stack repeat passed. Typed input is synthesized to audio and is half-duplex with microphone input. Session length and unified-memory headroom are explicit runtime limits.

The replay metadata field runtime.image records the local image label used when the exact calibration/evaluation tensors were captured. The label is provenance only, is not a downloadable dependency, and is not used by the conversion or runtime paths. Reproduction uses the pinned public source graph and the tensor/file hashes in the release manifest.

See the source repository's known-limitations and provenance documents for the complete operational record.

This source copy includes qualification completed after immutable weight revision a20c68547d27421c1e66bd501bf295569a0cc309 was published. It supersedes the model-card prose frozen at that revision; the qualified weight bytes and their signed manifest are unchanged.

License and attribution

The parent model and these redistributed derivatives are governed by OpenMDW-1.1; see LICENSE. NVIDIA origin and copyright notices are retained in THIRD_PARTY_NOTICES.md. This repository identifies the exact parent above and does not claim NVIDIA sponsorship or endorsement.