D0shi9/Barq-27B

Barq 27B | برق

Barq 27B is a local GGUF distribution with a Barq operating policy and a chat-template adapter embedded in the language model. It is derived from Prism ML's Ternary Bonsai 2 27B GGUF, whose upstream model card identifies Qwen3.8-27B as its backbone.

The numerical tensor weights are unchanged. Barq does not claim additional training, a new weight quantization method, or an improvement over the upstream model's benchmark results. Its changes are in the operating policy, chat template, and descriptive GGUF metadata.

Files

File	Purpose	Approximate download size
Barq-27B-PQ2_0.gguf	Language model with the embedded Barq policy and template	7.21 GB
Barq-27B-mmproj-Q8_0.gguf	Optional vision projector for image input in a compatible runtime	0.63 GB
USAGE.md	Runtime requirements and local serving examples	Text
LICENSE	Apache License 2.0	Text
NOTICE	Attribution and modification notice	Text
SHA256SUMS.txt	Integrity checksums for the other files	Text

Download sizes use decimal GB and are not RAM or VRAM requirements. Text-only use does not require the projector. Runtime memory also depends on the context, cache configuration, backend, and other allocations.

Runtime compatibility

Use Prism ML's modified llama.cpp runtime with support for the PQ2_0 ternary format and its activation transforms. The upstream model card states that stock llama.cpp cannot run these files correctly. Generic GGUF support alone is insufficient.

See USAGE.md and the upstream Bonsai demo for setup. Runtime binaries are obtained separately.

Intended use and application responsibilities

The embedded policy is intended for conversation, explanations, coding assistance, and planning. Arabic and English are target interaction languages; this statement is not a language-quality benchmark.

Tool calls require a host application that supplies tool definitions, validates requests, executes permitted actions, and returns results. A generated tool call does not itself run a command or access a device or account.

Persistent memory, MCP connections, agent orchestration, and the six named profiles with enforced token budgets are application features. A host must implement them; they are not supplied as services by these GGUF files. References to host features in the embedded policy apply only when that host actually provides them.

The host should preserve the embedded chat template. Replacing it or bypassing it changes how the Barq policy is applied. Prompts are guidance, not an authorization or security boundary.

Evaluation and limitations

No Barq-specific benchmark scores are published in this repository. Upstream scores are not presented as measurements of this distribution. No comparative speed, accuracy, reliability, or hardware-fit guarantee is made.

Model outputs can be incorrect, including generated code and image interpretations. Review outputs and validate proposed actions before relying on them. Image input requires the supplied projector and a compatible multimodal host. Compatibility with every GGUF application is not implied.

License and attribution

Distributed under the Apache License 2.0. See NOTICE for attribution and the scope of modifications. Original model authorship is retained; Barq does not imply endorsement or affiliation with upstream vendors.

العربية

برق 27B نسخة GGUF محلية مشتقة من النموذج الأصلي لـ Prism ML، تتضمن سياسة تشغيل وقالب محادثة خاصين ببرق. الأوزان العددية لم تتغير، ولا يُدّعى تدريب إضافي أو تفوق مقاس على النموذج الأصلي.

يلزم محرك Prism ML المتوافق مع PQ2_0. ملف الصور اختياري للمحادثة النصية. تنفيذ الأدوات والذاكرة الدائمة وMCP والوكلاء يحتاج تطبيقاً مضيفاً يوفر هذه الوظائف فعلياً. تفاصيل التشغيل في USAGE.md.