M2M-100 418M — GGUF (ggml)
GGUF / ggml conversion of facebook/m2m100_418M for use with CrispStrobe/CrispASR.
M2M-100 is a multilingual encoder-decoder transformer for machine translation between 100 languages — any direction, no English pivot required. The 418M variant has 12 encoder + 12 decoder layers (d=1024, 16 heads, FFN=4096, ReLU). Distributed under MIT license.
Files
| File | Size | Notes |
|---|---|---|
m2m100-418m-f16.gguf | 935 MB | F16 weights (reference quality) |
m2m100-418m-q8_0.gguf | 502 MB | Q8_0 quantized (identical quality to F16 on test set) |
m2m100-418m-q4_k.gguf | 272 MB | Q4_K quantized (minor word choice differences) |
Quick start
# 1. Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF
cmake --build build -j
# 2. Pull model
huggingface-cli download cstr/m2m100-418m-GGUF m2m100-418m-q8_0.gguf --local-dir .
# 3. Translate (standalone)
./build/bin/crispasr --backend m2m100 -m m2m100-418m-q8_0.gguf \
--translate "Hello world, how are you today?" \
--source-lang en --target-lang de
# 4. ASR + translate pipeline
./build/bin/crispasr -m ggml-base.en.bin -f samples/jfk.wav \
--translate-model m2m100-418m-q8_0.gguf --target-lang de
Quality verification
All three quant levels produce correct translations:
| Input (English) | F16 | Q8_0 | Q4_K |
|---|---|---|---|
| Hello world, how are you today? | Hallo Welt, wie bist du heute? | Hallo Welt, wie bist du heute? | Hallo Welt, wie bist du heute? |
| The president said he would not attend the meeting. | Le président a dit qu'il ne sera pas présent à la réunion. | ...qu'il ne participerait pas... | ...qu'il ne voulait pas assister... |
| Machine learning is changing the world. | El aprendizaje de máquina está cambiando el mundo. | (identical) | La aprendizaje... (minor) |
Supported languages (100)
af, am, ar, ast, az, ba, be, bg, bn, br, bs, ca, ceb, cs, cy, da, de, el, en, es, et, fa, ff, fi, fr, fy, ga, gd, gl, gu, ha, he, hi, hr, ht, hu, hy, id, ig, ilo, is, it, ja, jv, ka, kk, km, kn, ko, lb, lg, ln, lo, lt, lv, mg, mk, ml, mn, mr, ms, my, ne, nl, no, ns, oc, or, pa, pl, ps, pt, ro, ru, sd, si, sk, sl, so, sq, sr, ss, su, sv, sw, ta, th, tl, tn, tr, uk, ur, uz, vi, wo, xh, yi, yo, zh, zu
Architecture
Text → SentencePiece BPE tokenizer (128K vocab, 100 lang codes)
→ Source lang token (__en__) + text tokens + </s>
→ 12-layer transformer encoder (d=1024, 16 heads, FFN=4096, ReLU, pre-norm)
→ Sinusoidal positional embeddings (pre-computed)
→ 12-layer transformer decoder (self-attn + cross-attn + FFN)
→ Shared embedding LM head (tied weights)
→ Target lang forced as first decoder token
→ Greedy decode → translated text
Shared embedding: encoder, decoder, and LM head all use the same 128112×1024 table.
Conversion
python models/convert-m2m100-to-gguf.py \
--input facebook/m2m100_418M \
--output m2m100-418m-f16.gguf
Also supports facebook/m2m100_1.2B and facebook/wmt21-dense-24-wide-en-x (same architecture, different scale).
Related models
facebook/m2m100_418M— original PyTorch modelfacebook/m2m100_1.2B— larger 1.2B variantfacebook/wmt21-dense-24-wide-en-x— WMT21 competition winner (4.7B)
Provenance and EU AI Act Art. 53 note
- Upstream model: facebook/m2m100_418M — published by
facebook. - 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.