All stories
OrcaCyber Nano 1.0 Momentum Community reports

OrcaCyber Nano packs a 27B security model into a 13GB GGUF

OrcaRouter's local security model offers built-in MTP drafting and a 16GB-card example, with separate downloads for vision and DFlash speculation.

Sources (2)

Why now The model appeared in our October 27 Hugging Face trending discovery at 05:53 UTC. Its repository was created October 16; this is a current deployment recommendation, not a new-release claim or evidence of measured growth.

OrcaCyber Nano 1.0 offers a concrete option for security engineers who want to keep code, logs and disassembly on their own machine. OrcaRouter describes it as a 27B vision-language reasoning model for defensive work, distributed in GGUF format for llama.cpp. Its appearance in our October 27 Hugging Face trending discovery makes the deployment recipe worth a look, but the repository dates to October 16: this is not a launch today.

The smaller main file is 12.99GB. The author pairs it with a 16GB graphics-card example using a single slot, a 32,768-token context and the built-in MTP speculative drafter. A 15.70GB main build is also offered. The card’s larger-card example combines that build with a separate 2.06GB DFlash2 drafter and a 0.93GB vision projector. Image input requires the projector; the two main builds already contain their MTP heads.

These file sizes are not guarantees that every advertised feature or maximum context fits in 16GB. The card lists a 262,144-token model context, substantially above its example’s setting. Hardware recommendations and compatibility are the author’s guidance, not a deployment independently tested by AITrending.

The public English card can be read without signing in, but downloading files requires accepting the repository’s access conditions. Metadata lists a custom Qwen license rather than a standard permissive open-source license. OrcaRouter also says its benchmark harness, settings and full result tables are still being finalized, so the recommendation rests on the documented local workflow rather than treating its performance table as independently verified.