Full-precision (bf16) MLX-format conversion of empero-ai/Qwythos-9B-Claude-Mythos-5-1M, a Qwen3.5-9B–based multimodal (vision + text) reasoning model with a 1M-context configuration.

This is the unquantized MLX variant — weights are unchanged from the source, just re-saved in MLX's native layout with a sanitised config. For a smaller download see the 4-bit version.

Details


Base	`empero-ai/Qwythos-9B-Claude-Mythos-5-1M` (`Qwen3_5ForConditionalGeneration`, VLM)
Precision	bfloat16 (no quantization)
Size	~18 GB
Tool	`mlx_vlm.convert --dtype bfloat16` (mlx-vlm 0.4.4)
Architecture	`qwen3_5` VLM (qwen3_vl-style vision tower + linear-attention text model)

Why bf16?

This build is for maximum fidelity on Apple Silicon: the weights are bit-for-bit the original's, so output quality is identical — but it runs on MLX's Metal backend, which generates ~2× faster than the original on PyTorch-MPS at the same memory footprint. Pick this when you want the original's exact behaviour with a native Apple Silicon speedup, and you have the RAM to spare.

vs the original

Everything below is on the same Apple M4 Pro: identical 104-token chat prompt, 128 new tokens, greedy. "Original (PyTorch-MPS)" = the upstream empero-ai bf16 weights run through 🤗 transformers with its default attention path (no flash-linear-attention — the typical out-of-the-box experience). Perplexity is on a fixed 252-token English passage.

Metric	Original (PyTorch-MPS)	This build — MLX bf16
Generation	7.4 tok/s	14.5 tok/s (~2×)
Prefill	235 tok/s	270 tok/s
Peak memory	19.1 GB	19.1 GB (same)
Perplexity	5.74	5.74 (identical)

Bottom line: same weights, same quality, ~2× faster generation than the original's PyTorch-MPS path — pure framework gain, no quantization tradeoff. If you want a ~70%-smaller download and ~3× less memory (at ~25% higher perplexity), see the 4-bit build.

Usage

bash

pip install mlx-vlm
python -m mlx_vlm generate \
  --model BRlin/Qwythos-9B-Claude-Mythos-5-1M-MLX-bf16 \
  --prompt "Who are you?" --max-tokens 128

Also runnable with oMLX as an OpenAI-compatible chat model. Peak memory for text generation is ~19 GB.

Conversion notes

The source config needs two sanitisations before mlx_vlm.convert (mlx-vlm 0.4.x) accepts this model; both are baked into this repo's config.json:

text_config.rope_parameters.partial_rotary_factor — the YaRN-edited source moved this key to the top level of text_config; mlx-vlm reads it from inside rope_parameters. Re-added as 0.25.
vision_config.model_type — source uses HF's qwen3_5_vision; mlx-vlm's VisionModel only accepts qwen3_vl / qwen3_5 / qwen3_5_moe. Set to qwen3_5.

torchvision must also be importable in the conversion env (transformers' AutoProcessor builds Qwen3VLVideoProcessor). mlx-vlm's qwen3_5 rope does not apply YaRN scaling, so the YaRN keys in rope_parameters are inert; max_position_embeddings is left at 1,048,576.

Attribution

All model behaviour, identity, and training credit belong to Empero AI (the empero-ai/Qwythos-9B-Claude-Mythos-5-1M authors). This repo only provides an MLX-format conversion for Apple Silicon.