Modelpublic

andreribeiro87/Qwen3.5-35B-A3B-telco-tracka-GGUF

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

0likes82downloads

Model Card

Qwen3.5-35B-A3B — Telco Track A GGUF

GGUF quantizations of a distilled Qwen3.5-35B-A3B fine-tuned on 5G NR drive-test troubleshooting trajectories (Zindi Telco Troubleshooting Agentic Challenge — Track A).

The LoRA adapter was merged into the base weights before quantization.

Files

File	Quantization	Size	Notes
`Qwen3.5-35B-A3B.Q4_K_M.gguf`	Q4KM	~20 GB	Recommended — good quality/size trade-off
`Qwen3.5-35B-A3B.Q5_K_M.gguf`	Q5KM	~17 GB	Higher quality
`Qwen3.5-35B-A3B.BF16-*.gguf`	BF16 (sharded)	~66 GB	Full precision — re-quantize yourself
`*-mmproj.gguf`	BF16	~861 MB	Multimodal projection (Unsloth artefact)

Model Details

Property	Value
Base model	unsloth/Qwen3.5-35B-A3B
Architecture	Qwen3.5 MoE — 35B total params, ~3B active
Distillation teacher	DeepSeek-V4-Flash
LoRA config (pre-merge)	r=16, alpha=32, BF16, targets: q/k/v/o_proj
Task	5G NR drive-test fault diagnosis (multi-choice)
Quantization tool	Unsloth `save_pretrained_gguf`

How to Run (llama.cpp)

bash

# Q4_K_M — ~24 GB VRAM or RAM
llama-cli -m Qwen3.5-35B-A3B.Q4_K_M.gguf \
    --chat-template qwen3 \
    -p "You are a 5G NR troubleshooting expert..." \
    -n 512

# Q5_K_M
llama-cli -m Qwen3.5-35B-A3B.Q5_K_M.gguf \
    --chat-template qwen3 \
    -n 512

LoRA Adapter

The raw LoRA adapter (without merging) is available at: andreribeiro87/Qwen3.5-35B-A3B-telco-tracka-lora