@/jCSrSSSSSSSSS S . S S S SSSSSSS . SSSSSSSSSS . SSSSSSS S!S"S . S#S$S%S&S'SS(SS)S . S*S+S,S-S.SS/S0S1S . S2S3S4S5S6SS7S0S8S . S9S:S;SS0S?S . S@SASBSCSDSSES0SFS . / rgG)HuCShared model definitions for Hermes local model management. Single source of truth — both auto-model-switch.py and model-health-watchdog.py import from here. Add/edit models in ONE place. LIVE PATH: ~/.hermes/scripts/models.py (systemd + cron reference this path) This copy is canonical for reference. ⚠️ If this file has Q4_K_M quants in the model names BELOW, it is STALE. The live copy at ~/.hermes/scripts/models.py has the correct quant names (Q4_K_S, Q6_K, etc.) and includes all models (Qwen3-8B, Qwen3-4B-2507, Qwen3-14B, Gemma QAT). Run diff to sync if needed. z Qwen3.6-35BuB🌟 BEST DAILY DRIVER: biggest brain, reasoning, coding, creativeuMoE 35B (3B active/tok) — UD Q4_K_S from tvall43. Big gun for complex reasoning, multi-file refactors, long creative tasks. ~20GB at Q4 so needs --n-cpu-moe offloading on 11GB VRAM.zllama-server.serviceiz custom:localzQwen3.6-35B-A3B-UD-Q4_K_S.ggufi35b) namedescdetailserviceportprovider model_namecontextshortz Qwen3.5-9BuE🔓 UNCENSORED 9B: abliterated, no refusals, general chat & creativezy9B dense HauhauCS Aggressive Q6_K. Popular abliterated finetune, 0 refusals. Fits entirely on GPU. Must use --jinja flag.zllama-server-9b.serviceiz3Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q6_K.ggufi9bzQwen3-8Bu>⚡ QWEN3-8B: fast dense 8B, solid all-rounder for quick tasksz?8B dense Qwen3 UD Q6_K_XL. ~50-80 t/s on 1080 Ti. Full GPU fit.zllama-server-qwen8b.serviceizQwen3-8B-UD-Q6_K_XL.ggufi8bz SmolLM3-3Bu<🏎️ SMOLM3-3B: CPU-only, tiny/fast, for Hindsight memoryux3B SmolLM Q4_K_M — CPU-only (-ngl 0) on port 8087 for Hindsight memory. Do NOT switch this service via model-switcher.zllama-server-smol.serviceiz$HuggingFaceTB_SmolLM3-3B-Q4_K_M.ggufi@smolz Qwen3-14BuE⚡ QWEN3-14B: dense 14B, fast ~25 t/s, solid all-rounder replacementzV14B dense Qwen3 UD Q4_K_XL. Full GPU fit. ~25 t/s gen, ~135-200 t/s prompt processing.zllama-server-qwen14b.serviceizQwen3-14B-UD-Q4_K_XL.gguf14bz Qwen3-4B-2507uA⚡ QWEN3-4B-2507: newest tiny Qwen, Q8 for quality, full GPU fitzI4B Qwen3-4B-Instruct-2507 UD Q8_K_XL. Full GPU fit. Fastest local option.zllama-server-qwen4b.serviceiz&Qwen3-4B-Instruct-2507-UD-Q8_K_XL.ggufi4bzGemma-4-26B-A4Bu;🚀 GEMMA 4: Google's MoE, native vision, 256K ctx, 24 t/szMoE 26B (8 active/tok) UD-Q4_K_XL. Native 256K ctx (set to 128K). Full vision support via mmproj on CPU. ~24 t/s gen, ~40-55 t/s prompt. --n-cpu-moe 128. ~10GB/11GB VRAM.zllama-server-gemma4.serviceiz"gemma-4-26B-A4B-it-UD-Q4_K_XL.ggufgemma4z Gemma-3-12BuF🌟 GEMMA 3 12B: Google's 12B, native 128K ctx, UD uncensored variantz12B dense Gemma 3 UD-Q5_K_XL. Full GPU (-ngl 99, ~7.7GB VRAM). ~20 t/s gen, ~120-150 t/s prompt. Native 128K ctx. Turbo4 KV. Text-only GGUF.zllama-server-gemma3.serviceizgemma-3-12b-it-UD-Q5_K_XL.ggufgemma3zGemma-3-12B-QATu<🔬 GEMMA 3 12B QAT: quantization-aware-trained bonus modelz;12B gemma-3-12b-it QAT Q4_0 on port 8091. Bonus/side model.zllama-server-gemma3-qat.serviceizgemma-3-12b-it-qat-Q4_0.gguf gemma3qatN)__doc__MODELS'/home/rurouni/.hermes/scripts/models.pyrsT  TL)"6 WN,"K PS0"0 NM."< Wj1"1  S]0"> "M0": Xa0"6 "NO4"4 sd r