NEWS → · Open weights · Salesforce
Salesforce's SOV-Transcribe-1 is the new #1 open-weight speech model — and it needs a big-memory card
Nov 13, 2026
Salesforce released SOV-Transcribe-1 on Hugging Face on Nov 12, 2026 under Apache 2.0: a 26.6-billion-parameter speech model (4.6 billion active) that ranks #1 among open-weight models for non-streaming speech-to-text on Artificial Analysis, at a 2.28% word error rate. Its BF16 checkpoint is about 53 GB, so running it yourself starts on an 80 GB-class GPU.
Image: Salesforce via huggingface.co
From Salesforce's own announcement on Hugging Face (Nov 12, 2026), read first-hand: SOV-Transcribe-1 pairs a Whisper large-v3 encoder with a Gemma 4 MoE decoder, is non-streaming only, covers seven languages, and was tested on audio up to 30 minutes. Its 2.28% AA-WER beats NVIDIA's Nemotron 3.5 Transcribe (2.81%) and Mistral's Voxtral Small (2.77%) on Artificial Analysis's index, where it ranks #1 open-weight and #7 overall (rankings as of Nov 12, 2026). It also leads open-weight models on BeSimple's VoiceCodeBench entity test — recovering 93.4% of key entities such as emails and account numbers — and supports instruction-based style steering and keyword context biasing. Serving is via a vLLM plugin or Transformers; the technical report is still to come.
The hardware maths, from the model's own card: the BF16 checkpoint is about 53 GB, which does not fit a 48 GB card. On this board the cheapest verified way in is a single A100 80GB — Thunder Compute at $1.09 an hour, checked today — with the RTX PRO 6000's 96 GB the roomier single-card option, from $1.80 an hour at Nebius.
What it means if you rent GPUs
Transcription is one of the clearest sovereign-AI use cases — call recordings and dictation that cannot leave your own infrastructure — and the best open option now needs datacenter-class memory rather than a workstation card. That is a rental-shaped problem: an A100 80GB at about a dollar an hour, switched on only when there is audio to process, against closed per-minute APIs. The model is in our catalogue from today, with the GPU maths on its page.