Threads (@akiraxtwo)
local-llmai-tools
Original source

Qwythos-9B-Claude-Mythos-5 — 1M Context Open-Source Reasoning Model

Summary

Empero released Qwythos-9B-Claude-Mythos-5-1M, a full-parameter fine-tune based on a deeply uncensored Qwen3.5-9B.

Key specs:

  • Context: 1,048,576 tokens (1M) via YaRN
  • Training data: 500M+ tokens of Claude Mythos/Fable synthetic CoT data
  • Full-parameter fine-tune (not LoRA)
  • Benchmarks vs base Qwen3.5-9B: MMLU +34, gsm8k-strict +30

Hardware requirements (inference):

  • Q4KM quant: ~5.3GB VRAM (6-8GB GPU minimum)
  • Q8_0 quant: ~8.9GB VRAM
  • BF16 full precision: ~17GB VRAM
  • Full 1M context: 24GB+ VRAM needed for KV cache (otherwise CPU offload or reduce context)

Discussion notes:

  • One commenter questioned if this violates ToS (using “Mythos” access / Claude synthetic data)
  • Rare combination of long-context + high-quality reasoning in open-source 9B class