---

license: mit

language:

• zh
• en
• yue

pipeline_tag: automatic-speech-recognition

tags:

• audio
• speech-recognition
• transcription
• gguf
• glm
• zhipu
• multilingual

library_name: ggml

base_model: zai-org/GLM-ASR-Nano-2512

---

GLM-ASR-Nano-2512 — GGUF

GGUF conversions and quantisations of zai-org/GLM-ASR-Nano-2512 for use with CrispStrobe/CrispASR.

Available variants

| File | Quant | Size | Notes |

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

| glm-asr-nano.gguf | F16 | 4.3 GB | Full precision |

| glm-asr-nano-q8_0.gguf | Q8_0 | 2.3 GB | High quality |

| glm-asr-nano-q4_k.gguf | Q4_K | 1.3 GB | Best size/quality tradeoff |

All variants produce correct transcription on test audio.

2026-07 update — BPE merges baked in + long-form single-pass

All files were re-published with the tokenizer's BPE merges in the GGUF

metadata (tokenizer.ggml.merges, +2 MB). CrispASR ≥ this date uses them to

encode the transcription prompt exactly like the HF blueprint — earlier

GGUF+runtime combinations silently sent no instruction at all, which is

what caused repetition loops on noisy audio and empty output on long clips

(CrispASR #218).

Old GGUFs still work with the new runtime (it falls back to a baked default

prompt), but custom --ask / --language instructions need these files.

Long audio: --chunk-seconds 0 now decodes up to 655 s in one pass

(30 s encoder windows, one LLM prompt — the blueprint's layout), matching

the transformers reference verbatim on the #218 test clip. Note the model

skips leading non-speech audio in single-pass mode (blueprint behaviour);

the default 30 s-chunked mode covers more of such clips.

Model details

• Architecture: Whisper encoder (1280d, 32L, partial RoPE) + 4-frame projector + Llama LLM (2048d, 28L, GQA 16/4)
• Parameters: 1.5B
• Languages: Mandarin (+ Chinese dialects), English, Cantonese (model card metadata declares en, zh; prose adds Cantonese 粤语 + other Chinese dialects). Not a general multilingual model — no Japanese/Korean/European-language support.
• License: MIT
• Outperforms OpenAI Whisper V3 on benchmarks (lowest avg error rate 4.10)

Usage with CrispASR

git clone https://github.com/CrispStrobe/CrispASR && cd CrispASR
cmake -S . -B build && cmake --build build -j8

# Auto-detect backend from GGUF
./build/bin/crispasr -m glm-asr-nano-q4_k.gguf -f audio.wav

# Explicit backend
./build/bin/crispasr --backend glm-asr -m glm-asr-nano-q4_k.gguf -f audio.wav -osrt

Conversion

python models/convert-glm-asr-to-gguf.py --input zai-org/GLM-ASR-Nano-2512 --output glm-asr-nano.gguf
crispasr-quantize glm-asr-nano.gguf glm-asr-nano-q4_k.gguf q4_k

Provenance and EU AI Act Art. 53 note

• Upstream model: zai-org/GLM-ASR-Nano-2512 — published by zai-org.
• Upstream licence: mit. 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/GGML). 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.