SOV-Transcribe-1: Salesforce's Open-Weight Speech Model

Salesforce's non-streaming speech model targets transcription and entity accuracy. Model size, license documentation and what to test before deploying.

Salesforce released SOV-Transcribe-1 on November 12, 2026, for non-streaming speech recognition. Its launch emphasizes names, phone numbers and email addresses alongside ordinary transcription accuracy. This November 16 assessment examines recorded-audio use and deployment requirements; S5 Labs has not run the model. Salesforce announcement

A useful transcript must preserve the details a person will act on. A mistaken account number can matter more than several missing filler words. That makes entity accuracy worth testing separately from the overall proportion of words transcribed incorrectly.

Recorded audio, with instructions

The model card lists seven languages, audio inputs tested up to 30 minutes, and text output. Instructions can request verbatim or cleaned-up transcription; context can supply relevant vocabulary. This is a non-streaming ASR model, so it should not be priced or evaluated as a complete conversational voice agent. Our comparison of voice-agent APIs covers that broader service boundary.

Salesforce reports 76.3% task success on VoiceCodeBench, where a recording passes only when every target entity is correct, and 93.4% entity recovery on its companion metric. Those figures describe different scoring rules. They do not mean that 93.4% of calls have fully correct transcripts. The figures are Salesforce’s November 12 report, not our independent benchmark results. Launch evaluation

The full model determines memory needs

The card gives 26.6B total parameters and 4.6B active, including the audio encoder, and BF16 precision with an FP32 MoE router. Its components include a Whisper large-v3 encoder and a Gemma 4 26B A4B decoder. Transformers support uses repository modeling code; the documented vLLM path uses a plugin while native support is being upstreamed. Architecture and serving instructions

The repository listing shows approximately 53.3 GB and sixteen weight shards. That download size is not a measured runtime-memory requirement. Budget for the full checkpoint, audio processing, caches and concurrency; active parameter count describes computation, not a small standalone download. We have no tested GPU configuration to recommend for this model.

What the license documentation establishes

Salesforce’s model card declares Apache 2.0, including in its license section. On November 16, the repository file listing contained no standalone LICENSE file, and fetching that path returned 404. We can report the declaration; we cannot say we inspected a separate license file for this checkpoint. Card · Files

Before deployment, retain the exact revision and its documentation, review the supplied code and component terms, and resolve any licensing uncertainty with the publisher. The presence of a permissive label alone is not our certification of every component or intended use.

A practical evaluation set

For a recorded-call workflow, start with audio you are authorized to process and a human-checked reference. Include difficult names, spoken identifiers, interruptions and the accents your users actually have. Score complete entity correctness alongside word error rate; also record omissions and words invented during silence.

Run the same recordings with and without vocabulary hints. Include plausible but absent terms in a separate test to see whether hints introduce errors. Keep verbatim and cleaned-up outputs separate, since removing a hesitation may be correct for one task and a failure for another.

These are proposed evaluation steps, not measured outcomes. A deployment decision should include processing time, peak memory and reviewer correction effort on those recordings. The launch’s accuracy figures help choose a candidate to test; your workload determines whether its errors are acceptable.

Reference guide

The Open Models guide compares open-weight models by license, hardware and task fit.

Continue reading.

Insight5 min read

Microsoft Phi-4-Reasoning-Vision: A Model That Knows When Not to Think

Phi-4 15B model skips structured reasoning for simple perception tasks, uses it only for math and science. Selective reasoning improves latency and cost.

Insight11 min read

Liquid LFM3: Six 2B-Class Open-Weight Models, One Revenue-Capped License

Liquid's LFM3 ships text, vision, audio, omni and two decision models near 2B parameters under the LFM Open License. Sizes, memory, runtimes, $10M line.

Insight11 min read

FLUX 3 Image Klein: 5.6B Open Weights That Target 8 GB GPUs, Free Below $5M Revenue

FLUX 3 Image Klein: 5.6B open weights for image generation and editing, an INT8 build pitched at 8 GB GPUs, free under $5M revenue, and a $0.008 API tier.