Two October 28 speech developments point to different deployment choices. NVIDIA released Nemotron 3.5 Transcribe, while IBM published an ONNX export of Granite Speech 5.0 2B for browser inference. Our coverage is dated October 29. The IBM files are a deployment update to an existing model, not a newly trained foundation model.
NVIDIA: tell the transcriber what kind of text you need
Nemotron supports instruction-driven transcription, translation, speaker attribution and audio questions. NVIDIA describes support for 40 languages and recordings up to an hour. Its exact checkpoint card lists roughly 34B total parameters including the audio encoder, with about 4.3B active. Active parameters should not be mistaken for the total weight footprint.
This is an offline audio model. Internal chunking does not turn it into live streaming recognition. The reference deployment uses an H100 80 GB; that example is not an RTX 3060 fit claim.
NVIDIA’s headline serving result uses its optimized Transcribe NIM. The card says those optimizations are not yet fully in the open-source release, and that release-day NeMo/vLLM integration is expected to be roughly twice as slow. That distinction matters before applying the advertised number to another runtime. Checkpoints use OpenMDW-1.1; NVFP4 targets Blackwell hardware. We have not reproduced the reported results.
IBM: a browser export with a substantial first download
IBM’s October 28 upload packages four-bit ONNX weights for Transformers.js and WebGPU. The three listed files total about 1.7 GB, before runtime and working memory. Its README requires Transformers.js 4.3 or newer for Safari 26 WebGPU support.
The Apache-2.0 export derives its weights from granite-speech-5.0-2b. The documented quantization check includes a small 73-utterance LibriSpeech sample; that is useful conversion evidence, not a broad independent reliability study.
For builders, the practical questions are runtime support, download size, language coverage and real recording quality. Neither artifact is an installed launcher feature, and we have not downloaded weights or run a transcription test.
Original AiBrainWorX reporting; sources checked October 29, 2026. Publisher claims and documented behavior are distinguished from independent testing, which was not performed.
