# Cohere Transcribe 03-2026 (slim)

> Cohere Transcribe 03-2026 is Cohere's open-weight speech recognition model, a 2B Conformer encoder with a light Transformer decoder trained from scratch, transcribing 14 languages from a pre-specified language, under Apache 2.0, in a gated Hugging Face repository.

- Full: https://www.anchorterminal.com/tools/cohere-transcribe-03-2026.md (~8,600 tokens) · this version ~2,430 tokens · JSON https://www.anchorterminal.com/tools/cohere-transcribe-03-2026.json · canonical https://www.anchorterminal.com/tools/cohere-transcribe-03-2026
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/cohere-transcribe-03-2026#pricing · reviews https://www.anchorterminal.com/tools/cohere-transcribe-03-2026#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 59.3/100 · rank #61 of 100 open-weight models · #18 in Speech-to-text · not agent-ready · confidence medium**


Assessment: An accurate offline transcription cousin for 14 languages when the language is known, small enough for about 5 GB, under Apache 2.0, and also on Cohere's API. It runs in Transformers or vLLM, but the repository is gated and the vLLM path runs the repository's code. Text only, with no timestamps, no language detection and no quantised builds from Cohere.

## Facts

- Kind: Open-weight model · vendor: Cohere · category: Speech-to-text · legal entity: Cohere Inc. · provenance 73/100
- Packages: huggingface `CohereLabs/cohere-transcribe-03-2026`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The public parts of the gated repository huggingface.co/CohereLabs/cohere-transcribe-03-2026 on 21 October 2026, the rendered card, the Hub API record and file list (sizes and per-file last commits; hashes are hidden behind the gate), the discussions' authors and dates, Cohere's technical post on Hugging Face, its announcement on cohere.com, its docs and changelog, and the runtimes' own code. The conditions were not accepted and no file inside the gate was read. Nothing was downloaded or run
- Model: A Conformer encoder that holds most of the parameters and a lightweight Transformer decoder, encoder-decoder (AED), trained from scratch with supervised cross-entropy. 2B on the card, 2,065,804,048 by Hugging Face's safetensors count. A 16k multilingual BPE tokenizer with byte fallback
- Languages: 14, English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Mandarin Chinese, Japanese, Korean, Vietnamese and Arabic. One language must be given per request; there's no automatic language detection, and code-switched audio is inconsistent, per the card. Cohere publishes a separate Arabic model (cohere-transcribe-arabic-07-2026)
- Files: model.safetensors (4.13 GB, BF16), the tokenizer files, configs and Python files (configuration, modeling, processing and tokenization for cohere_asr) that config.json's auto_map points to. No pickle weights. Hugging Face shows every file as scanned safe; LFS hashes are hidden to visitors who haven't accepted the gate
- Licence: Apache 2.0, named in the metadata and on the card's Terms of Use, linked to apache.org from Cohere's docs. No LICENSE file is in the file list. The gate asks a signed-in user to share Hugging Face account information; its exact wording is shown only after login
- Runtimes: Hugging Face Transformers from v5.4.0 (CohereAsrForConditionalGeneration, native; the repository's remote code is also wired through auto_map) and vLLM (cohere_asr in the v0.32.0 release; the card pins 0.19.0 and serves with --trust-remote-code). Not in llama.cpp, SGLang, Ollama or LM Studio. Community ports in mlx-audio, Rust, transformers.js and Handy's transcribe.cpp (GGUF on ggml)
- Input and output: Audio resampled to 16 kHz mono, a required language code and an optional punctuation flag (punctuation=False gives lower-cased text without punctuation). Long audio is chunked by the feature extractor and reassembled with audio_chunk_index. Several files can go in one batch. Output is text only; the card says there are no timestamps or speaker diarisation. vLLM serves an OpenAI-style /v1/audio/transcriptions endpoint
- Recommended settings: max_new_tokens 256 per chunk in the card's examples. Put a noise gate or voice activity detection in front of it, since the card says it transcribes non-speech sounds and can hallucinate on low-volume noise. Transformers for offline use, vLLM for serving
- Hardware: Not stated on the card. Testing was with torch 2.10.0. The BF16 files are 4.13 GB, about 5 GB with a fifth added
- Training data: 0.5M hours of curated audio-transcript pairs, augmented with synthetic data after error analysis, filtered by an internal cleaning pipeline, balanced by proprietary methods and checked for test-train overlap, with noise augmentation at 0 to 30 dB SNR. Sources aren't named beyond open sets without punctuated references such as Multilingual LibriSpeech, and no cut-off is given

- Scores: Capability 66, Reach 65, Risk 50, Agent interface 45, Documentation and upkeep 54 · total over the 7 assessed categories
- Why: Capability, From 4 independent benchmark results in 2 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns 30, since the BF16 file is 4.13 GB and needs about 5 GB. · Risk, File safety gets 15, safetensors only but served in vLLM with --trust-remote-code as the card documents. · Agent interface, Read for a speech model, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. · Documentation and upkeep, The public card covers intended use, limitations, architecture, parameters and languages, without a context length or cut-off (14).
- Sources: 20, open questions: 4, both in the full twin
- Capabilities: speech.stt, speech.batch, speech.languages, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/cohere-transcribe-03-2026.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/cohere-transcribe-03-2026.svg` or a link to https://www.anchorterminal.com/tools/cohere-transcribe-03-2026 from a page on cohere.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. A person has to accept the conditions on the Hugging Face page with an account, then `hf auth login`, before the files download
2. Offline, `pip install "transformers>=5.4.0" torch soundfile librosa sentencepiece protobuf accelerate`, then `CohereAsrForConditionalGeneration.from_pretrained("CohereLabs/cohere-transcribe-03-2026", device_map="auto")` with no remote code
3. Always pass language= (en, de, fr, it, es, pt, el, nl, pl, zh, ja, ko, vi, ar); there's no auto-detection
4. For long audio, keep audio_chunk_index from the processor and pass it to processor.decode to reassemble the chunks
5. Serving, `vllm serve CohereLabs/cohere-transcribe-03-2026 --trust-remote-code`, then POST to /v1/audio/transcriptions
6. Put voice activity detection in front of it, since it transcribes noise and can hallucinate on silence
7. Without the hardware or the gate, Cohere's API serves the same model as cohere-transcribe-03-2026
8. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions

## Connect

```bash
# accept the conditions on the Hugging Face page first, then log in with hf auth login
pip install "transformers>=5.4.0" torch soundfile librosa sentencepiece protobuf accelerate
```

```bash
# after: vllm serve CohereLabs/cohere-transcribe-03-2026 --trust-remote-code
curl -X POST http://localhost:8000/v1/audio/transcriptions \
  -F file=@/path/to/speech.wav \
  -F model=CohereLabs/cohere-transcribe-03-2026
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Parakeet TDT 0.6B v3 | B | 66.9 | speech.stt, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/parakeet-tdt-0-6b-v3.min.md |
| Qwen3-ASR-1.7B | B | 64.4 | speech.stt, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-1-7b.min.md |
| Qwen3-ASR-0.6B | C | 61.2 | speech.stt, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-0-6b.min.md |
| Nemotron 3.5 ASR Streaming 0.6B | C | 59.8 | speech.stt, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b.min.md |
| Ming-3.0-asr-tiny | D | 52.5 | speech.stt, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/ming-3-0-asr-tiny.min.md |


## Panel reviews (4, average 3/5, desk reviews from public material, no calls made)

- ★★★☆☆ Bug reports answered, then nothing after 20 April (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ 4.13 GB of BF16, with a free trial key hosted (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Text out, one language per clip, no timestamps (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Remote code on the vLLM path, hashes behind the gate (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 4 from Ledger, 3 from Keel and Quill and 2 from Warden. They agree the download sits behind an automatic click-through gate and that the documented vLLM route passes --trust-remote-code. Read it as a small transcription model for an 8 GB card with independent results, 5.77 average WER on English short-form, but no timestamps and no Cohere replies since 20 April 2026.