This repository provides a patched version of IsValorum/Cyber-Tiel-Coder-35B-A3B-APEX-I-NanoPlus-GGUF with nitayrabi's LoopFix weight-space and chat-template interventions applied.
By combining IsValorum's 4-moment Importance Matrix mixed-precision NanoPlus quantization (~11.5–12.3 GiB, tailored for 16 GB GPUs with generous context headroom) with nitayrabi's zero-fine-tuning LoopFix patch, this build prevents multi-turn reasoning loops and tool-call stalls during local agentic coding workflows.
APEX-I NanoPlusIQ2_S/IQ2_XXS in core feedforward blocks while keeping routers and attention layers at higher precision—leaving ample VRAM for 16k–32k context windows on 16 GB GPUs.LoopFixblk.26.ffn_down_shexp.weight and blk.34.ffn_down_shexp.weight), leaving early/mid syntax-binding layers untouched._preserve_thinking = false by default and tightens intra-turn retention and tool-use instructions so llama-server does not feed stale reasoning traces back into subsequent agent turns.llama.cpp)Run with llama-server using the recommended LoopFix sampling and reasoning parameters:
llama-server \
-m .\Cyber-Tiel-Coder-35B-A3B.APEX-I-NanoPlus-LoopFix.gguf \
--jinja \
-fa on \
-ngl 99 \
--top-p 0.95 \
--temp 0.6 \
--top-k 20 \
--min-p 0.05 \
--presence-penalty 0.0 \
--frequency_penalty 0.0 \
--repeat-penalty 1.0 \
--cache-type-k q8_0 \
--cache-type-v q8_0 \
--chat-template-file .\patched-chat-template.jinja \
--reasoning-effort xhigh
IsValorum/Cyber-Tiel-Coder-35B-A3B-APEX-I-NanoPlus-GGUF