| --- |
| language: |
| - hu |
| license: mit |
| library_name: mlx |
| tags: |
| - hungarian |
| - emese |
| - from-scratch |
| - llama |
| - base-model |
| - mlx |
| - apple-silicon |
| - quantized |
| pipeline_tag: text-generation |
| --- |
| |
| # Emese-Ér (517M) — MLX q8 |
|
|
| **MLX q8-quantized** build of Ér, the edge/embedded/research tier of the [Emese](https://emese.tech) |
| Hungarian model family — a from-scratch 517M-parameter foundation model (not EuroLLM-based). See the |
| `er/` repo's README for full architecture, training, and limitations details; this file only covers the |
| q8-specific notes. |
|
|
| | | | |
| |---|---| |
| | **Quantization** | q8, group size 64 (`mlx_lm convert -q --q-bits 8 --q-group-size 64`) | |
| | **Size on disk** | ~526 MB (vs. ~987 MB bf16) | |
| | **Quality** | ≈ bf16, near-lossless at 8-bit | |
| | **Max context length** | 32,768 tokens (unchanged from bf16 — quantization doesn't affect context) | |
|
|
| ## Usage |
|
|
| ```python |
| from mlx_lm import load, generate |
| model, tok = load("er-mlx") |
| print(generate(model, tok, prompt="A magyar nyelv", max_tokens=100)) |
| ``` |
|
|
| Base model only — no chat template. This checkpoint loads via `mlx_lm` despite its non-standard |
| `config.json` schema (see the `er/` README for why plain `transformers` can't load either variant). |
|
|
| ## Training |
|
|
| Identical underlying weights to `er/` (bf16), just quantized after the fact — no separate training. See |
| `er/README.md` for the full from-scratch pretraining details (~4.5B Hungarian tokens, no SFT/DPO). |
| ## Benchmarks |
|
|
| **Not applicable** — see `er/README.md`: Ér has no chat template or instruction-tuning, so the |
| chat-instruction benchmarks used for the other three tiers (`emese-bench`) don't apply. |
|
|
