CobrIX / CobrIX-1.5-preview-Coder-Full-72B-A18B-GGUF

PREVIEW 1.5 — GGUF quants for testing. Final release will follow after think-SFT, native MTP and updated quants.

CobrIX-1.5-preview-Coder-Full-72B-A18B-GGUF

GGUF quantizations of CobrIX/CobrIX-1.5-preview-Coder-Full-72B-A18B — MoE decoder (~72B total / ~18B active per token, 13 experts), DPO-aligned + SFT-reinforced for code and cybersecurity (PT/EN).

Quants

Q5_K_M = best quality here; Q4_K_M = smaller/faster with minimal loss. ChatML template embedded ( <|im_start|> / <|im_end|> ), stop at <|im_end|> . Needs a strong machine — on 16 GB VRAM use partial GPU offload.

Usage (Ollama)

FROM ./CobrIX-1.5-preview-Coder-Full-72B-A18B-Q5_K_M.gguf

TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""

PARAMETER stop "<|im_end|>"
PARAMETER num_ctx 8192

ollama create cobrix15-full-q5 -f Modelfile
ollama run cobrix15-full-q5 "Explique SQL injection e como mitigar."

Usage (llama.cpp)

llama-cli -m CobrIX-1.5-preview-Coder-Full-72B-A18B-Q5_K_M.gguf \
  --jinja -p "<|im_start|>user\nExplique SQL injection<|im_end|>\n<|im_start|>assistant\n" -n 300

Feedback — help make it better

Found a flaw or have an idea? 📧 [email protected] — send the prompt, the output, what you expected, and how to improve it.

Donations for Infrastructure

Support open-source CobrIX development:

Bitcoin (BTC)

bc1q8mu8fjak4y84qj4dlk8pu4d3zhknm92zra4r4m

Ethereum (ETH / ERC-20)

0x8D9187dEa0a77390ef668361cd5b236DE54af2BB

Solana (SOL)

GQR2jZnWuWP1c3dbuz4mC7ZnyveacBKy63q8qf9nj8bp

CobrIX — Open AI infrastructure and custom model research.