---

license: apache-2.0

language:

• ar

pipeline_tag: automatic-speech-recognition

tags:

• audio
• speech-recognition
• transcription
• gguf
• conformer
• arabic
• crispasr

base_model: CohereLabs/cohere-transcribe-arabic-07-2026

---

cohere-transcribe-arabic-07-2026 — GGUF

GGUF weights for CohereLabs/cohere-transcribe-arabic-07-2026 — Cohere Labs' 2B-parameter Arabic ASR model (a FastConformer encoder + Transformer decoder), released July 2026 (~11% WER on FLEURS Arabic per the source card).

These GGUFs run on CPU/Metal/CUDA/Vulkan via CrispASR — a C++ runtime for the Cohere Conformer-encoder / Transformer-decoder architecture.

> License: Apache 2.0 (inherited from source model). See the original model card for full terms.

---

Files

| File | Size | Type |

|------|------|------|

| cohere-transcribe-arabic-f16.gguf | 4.1 GB | F16 (reference precision) |

| cohere-transcribe-arabic-q8_0.gguf | 2.4 GB | Q8_0 |

| cohere-transcribe-arabic-q4_k.gguf | 1.5 GB | Q4_K |

| cohere-transcribe-arabic-q4_k-imatrix.gguf | 1.5 GB | Q4_K + importance matrix (Arabic-calibrated) |

| cohere-transcribe-arabic-ref.gguf | small | per-stage reference activations for crispasr-diff |

All quants keep LayerNorm weights / biases at F32. The -imatrix build is

calibrated on CC0 Common Voice Arabic (fsicoli/common_voice_17_0, ar/dev) and

recovers quality at the 4-bit budget; prefer it over plain q4_k.

On an M1 (Metal), end-to-end encode+decode on an 11 s clip is on par with the

transcribe.cpp reference runtime (~1.5 s, load excluded).

---

Quick start

# Build CrispASR (see the repo for full instructions)
git clone --recursive https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -G Ninja -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

# Shorthand: auto-downloads the recommended imatrix GGUF and defaults
# the language to Arabic — no -m / --hf-repo / -l needed.
build/bin/crispasr --backend cohere-ar audio.wav

cohere-ar is a CLI alias for the cohere backend: it routes to the same

runtime, resolves cohere-transcribe-arabic-q4_k-imatrix.gguf via -m auto,

and sets -l ar unless you pass an explicit -l (which always wins — useful

if you want to run this model with the LID pre-step or force another

language for testing). It's equivalent to:

build/bin/crispasr --backend cohere \
    --hf-repo cstr/cohere-transcribe-arabic-07-2026-GGUF:cohere-transcribe-arabic-q4_k-imatrix.gguf \
    audio.wav -l ar

Or point -m at a locally downloaded GGUF:

build/bin/crispasr --backend cohere -m cohere-transcribe-arabic-q4_k.gguf audio.wav -l ar

---

Architecture

| Component | Details |

|-----------|---------|

| Encoder | 48-layer FastConformer, d=1280, heads=8, head_dim=160, ffn=5120, conv_kernel=9 |

| Decoder | 8-layer causal Transformer, d=1024, heads=8, head_dim=128, ffn=4096 (ReLU), max_ctx=1024 |

| Vocab | 16,384 SentencePiece tokens |

| Audio | 16 kHz mono, 128 mel bins, n_fft=512, hop=160, win=400, 8× time subsampling |

| Parameters | ~2B |

Prompt tokens (decoder): the runtime prepends decoder_start_token_id (▁,

13764) to the control-token prompt, matching the reference decoder_input_ids.

Mel frame count is floor(n/hop)+1 (NeMo FilterbankFeatures), giving `T_enc =

calc_length(T_mel)`; both are required for the cross-attention context to line up.

---

Validation (crispasr-diff)

cohere-transcribe-arabic-ref.gguf holds per-stage reference activations dumped

from the transformers model by tools/dump_reference.py. Reproduce the

per-layer parity check with:

build/bin/crispasr-diff cohere cohere-transcribe-arabic-f16.gguf \
    cohere-transcribe-arabic-ref.gguf audio.wav

---

Related

• Source model: CohereLabs/cohere-transcribe-arabic-07-2026
• English sibling: cstr/cohere-transcribe-03-2026-GGUF — Cohere Transcribe 2B (lowest English WER)
• C++ runtime: CrispStrobe/CrispASR

Provenance and EU AI Act Art. 53 note

• Upstream model: CohereLabs/cohere-transcribe-arabic-07-2026 — published by CohereLabs.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.