Hy-MT1.5 1.8B MLX 2-bit — Lorqa conversion

Derived from the exact Tencent Hy-MT1.5 1.8B 1.25bit GGUF snapshot documented in LORQA_CONVERSION.json. This is a Lorqa conversion, not an official Tencent MLX release. No alternative publisher weights were used.

For Apple Silicon GPU inference with MLX-LM (hunyuan_v1_dense). Approximately 746 MB weights plus tokenizer. Main STQ linear weights are repacked exactly into affine 2-bit/group64. Tied embeddings are requantized from Q6_K to affine 8-bit/group64; norms are cast to FP16. The overall conversion is not lossless. No Python model code or custom Metal kernels are required for this affine version.

Conversion tooling: https://github.com/kuotient/hy-mt2-mlx at cc3b9ce816a3ea40aa470361ec71e43d4104d7a9. Used sherry_mlx.convert with the official Tencent HY-MT1.5 reference config/tokenizer and embed_bits=8.

from mlx_lm import load, generate
model, tokenizer = load("Lorqa/Hy-MT1.5-1.8B-MLX-2bit")
prompt = tokenizer.apply_chat_template([
    {"role": "user", "content": "Translate the following segment into Chinese, without additional explanation:\n\nHello, world!"}
], tokenize=False, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=256))

Use a user-only translation instruction; this translation-specialized model is not a general instruction/cleanup model. Preserve product names with terminology constraints and validate quality on your workload.

Validation

On M1 Max, Python MLX 0.32.2 / MLX-LM 0.31.3: three short/medium greedy outputs matched the original CPU GGUF. A 146-input-token paragraph measured approximately 213 output tokens/s and 0.73 s warm request time (load excluded). These are small standalone probes, not a broad accuracy, cold-start, Swift integration or concurrent-ASR benchmark. Runtime memory exceeds weight size.

License and attribution

The original Tencent HY Community License Agreement remains applicable; read License.txt and Notice.txt, including territory/use restrictions and redistribution requirements. This model is NOT Apache-2.0. Preserve both files when redistributing. LORQA_CONVERSION.json records the conversion and individual artifact hashes.