Qwythos-9B-v2 Fixes Repetition Loop Issue
techlatest_net · reddit · 2026-07-14
The Qwythos-9B-v2 GGUF card updates model usage instructions and improvements: the new version trains away the loop degradation issue under greedy or low-temperature decoding from the previous version, reducing loop behavior from 6.7% to 0%, while restoring the native MTP head and cleaning up the identity prompt; knowledge and reasoning capabilities remain at least as good as the base version.
The training fix method is called FTPO (Final-Token Preference Optimization), which focuses preference optimization only on the starting token that triggers the repetition loop, guiding the model toward coherent alternatives while minimally affecting the rest of the distribution. The post also provides loading instructions for llama.cpp, Ollama, LM Studio/jan/KoboldCpp, as well as a list of required files for normal text weights, MTP weights, and the vision projector for image input.
More from Models
- Kimi K3.1 Pro tops MathArena open models, closing frontier gap to about 6 months — teortaxesTex · 2026-10-26
- Ruff creator Charlie Marsh: frontier models find infinite bugs in any software — charliermarsh · 2026-10-26
- Dev testing one skill across Opus 5.5 and Sol 6.1 finds instruction following 'all over the map' — lucasmeijer · 2026-10-26
- supermemory unveils learner-1, claiming continual learning for any model — julianweisser · 2026-10-26
- Anthropic says Opus 5.5 cuts cost per task by ~40% vs Opus 5 — taufiqintech · 2026-10-26
- AI search APIs ranked: top 12 within 7 points, yet prices differ nearly 7x — FinanceYF5 · 2026-10-26