EngineeringThe briefing
AI-generated illustration
Nvidia launched an open-weight 30B mixture-of-experts transcription model that accepts plain-language prompts for context, speaker labels, translation, and audio questions. It supports 40 languages and runs on a single GPU via vLLM or NeMo-Speech.cpp.
Nvidia released Nemotron 3.5 Transcribe as a single checkpoint that combines transcription, speaker attribution for 13 languages, translation for 31 languages, and English audio Q&A. The 30B mixture-of-experts model uses 4 billion active parameters and handles recordings up to one hour. It is available today on Hugging Face under the OpenMDW-1.1 license and can be served with vLLM or run locally with NeMo-Speech.cpp on supported hardware from Jetson AGX Thor to Blackwell GPUs.
Users supply context via prompts such as medical terminology lists or formatting rules before transcription begins. Nvidia reports that fine-tuning on a medical dataset reduced word error rate on held-out speakers from 22.44 percent with prompting alone to 12.17 percent after 14 GPU-hours on a GB300. The company also released Granary v2, a data curation pipeline, and public manifests covering 463062 hours across 48 languages.
On independent benchmarks the model ranks first on Hugging Face Open ASR for English with 5.33 percent average word error rate. Artificial Analysis measured 2.81 percent word error rate, close to Voxtral Small at 2.77 percent while using roughly one-sixth the active parameters. The model does not support live streaming and was trained for recorded audio only.
Quantized versions for local RTX Spark and DGX Spark deployment are listed as coming soon. Developers can fine-tune the open weights with NeMo Speech tools to adapt to domain vocabulary.
Builders evaluating transcription pipelines can test Nemotron 3.5 Transcribe on Hugging Face to measure prompt-driven error reduction against their current stack. An illustrative scenario is a developer preparing a medical-meeting workflow: supply a prompt listing domain terms and requesting speaker labels, then compare the resulting transcript against a baseline model on the same audio files. The evidence shows a 45.8 percent relative error reduction after fine-tuning, yet does not report runtime memory or latency on consumer GPUs. Unknown factors include exact token throughput on RTX hardware and whether the 4B active parameters deliver lower latency than larger dense models under identical batch sizes.
Operators considering domain-specific transcription should first verify data residency and retention terms under the OpenMDW-1.1 license before uploading recordings. An illustrative scenario is a healthcare team that fine-tunes the model on internal notes using the released Granary pipeline on a single GB300 for 14 GPU-hours, then routes production audio through the adapted checkpoint. The supplied evidence establishes the word-error improvement on held-out speakers but does not state whether fine-tuned weights may be redistributed or whether Nvidia retains any usage telemetry from the fine-tuning run.
Download the model weights from Hugging Face and run a 10-file test set of representative audio through both prompted and fine-tuned versions while recording word error rate, speaker attribution accuracy, and GPU memory used. Compare the measured values against the current production transcription service to decide whether the prompt interface and parameter count justify migration.
Source material behind this briefing. Our practical implications are Ashlr’s interpretation.
This briefing was prepared with AI assistance from the linked sources. The analysis may include interpretation; consult the original sources for complete context. Cover artwork is illustrative.
GitHub released Copilot Fusion, a generally available system that selects and orchestrates models for coding tasks. Enterprise admins can restrict the model pool while developers choose effort levels instead of individual models.
EngineeringThe briefing
AI-generated illustration
Merged changes across Phantom Secrets, Phantom, and Locus include dry-run modes for vault commands, clearer doctor messages, and credential routing updates. Several reworks resolved merge conflicts on main.
EngineeringAshlr release
AI-generated illustration
Ashlr MD now drops deleted notes from activity lists and restores sessions without overwriting tabs. Other merges add contract tests, memory bounds and routing accounting across products.
