Qwythos-9B-Claude-Mythos-5-MTPLX

Qwythos-9B-Claude-Mythos-5-MTPLX

An MTPLX-ready build of empero-ai/Qwythos-9B-Claude-Mythos-5-1M with a working multi-token-prediction (MTP) head, for faster local decoding on Apple Silicon with exact rejection-sampling.

• Body: converted to MLX and quantized to 4-bit (affine, group size 64).
• MTP head: kept in bf16 (small but drives draft acceptance).
• On disk: ~5.2 GB.

Why this exists

The original empero-ai/Qwythos-9B-Claude-Mythos-5-1M safetensors release ships without an MTP head — only the companion ...-1M-GGUF repo carries a restored Qwen3.5-compatible MTP head (for llama.cpp --spec-type draft-mtp). MTPLX runs MLX, not GGUF, and its Forge tool only carries over an existing MTP head rather than training one.

This build bridges the gap: the bf16 MTP head was extracted from the GGUF's Qwythos-9B-Claude-Mythos-5-1M-MTP-BF16.gguf (blk.32 nextn tensors), remapped to MLX/MTPLX key names, paired with the full-precision body, and run through mtplx forge build. MTP-contract calibration reached exact agreement (1.0) at depths 1–3, confirming the head matches the trunk.

Architecture

Usage (MTPLX)

mtplx serve --model <this-model-dir> \
  --chat-template-profile tokenizer \
  --reasoning-parser qwen3

Notes:

• This is a reasoning model — it thinks before answering (<think> block), so allow a generous max_tokens (>=2048).
• Use --chat-template-profile tokenizer so the model's own chat template (with <think> handling) is used, and --reasoning-parser qwen3 to fold the thinking segment.
• Default sampling: temperature=0.6, top_p=0.95, top_k=20.

Provenance & license

• Body: empero-ai/Qwythos-9B-Claude-Mythos-5-1M
• MTP head: empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF
• Forged locally with MTPLX Forge (see mtplx_runtime.json for the contract and verification details).
• License: Apache-2.0, inherited from the base model. All credit for the base model and the restored MTP head belongs to Empero AI.