---

license: apache-2.0

base_model:

- Qwen/Qwen3-8B

tags:

- gguf

- quantized

- uncensored

- on-device

- opal

- qwen3

pipeline_tag: text-generation

language:

- en

---

Cognis Opal — Base Weights (GGUF)

**Recommended base weights for the Cognis Opal

on-device assistant system.** One file, drop it in and run.

> Attribution — read this. These weights are a GGUF quant of OmniCoder-Qwen3.5-9B, an

> uncensored community model built on the Qwen3.5 (Qwen3) base by Alibaba (Apache-2.0). They are

> rehosted here for convenience so Opal runs in one step. **Cognis Digital did NOT train these

> weights. Cognis's contribution is the Opal system** — Aleph O(N) memory + retrieval +

> test-time compute + multimodal orchestration — which wraps any open model. Full credit to the

> Qwen team and the OmniCoder quant author.

What it is

• Architecture: Qwen3.5, ~9B params
• Format: GGUF (single file), CPU/GPU via llama.cpp / Ollama / LM Studio
• Trait: uncensored community fine-tune; pristine — Opal never modifies the weights

License

Apache-2.0 (inherited from the Qwen3.5 base). Redistributed with attribution per the license. Cite

the base model and quant author if you use it. See the base model cards for full terms.

Use with Opal

# with Opal's system (memory + reasoning + multimodal):
git clone https://huggingface.co/cognis-digital/cognis-opal
# put this .gguf in cognis-opal/model/  (or pull the Ollama tag), then:
python -m synthos chat --adaptive

Or standalone via llama.cpp:

llama-cli -m cognis-opal-9b.gguf -p "Hello"

Honest scope

A ~9B model is not a frontier model and won't beat GPT-5.5 / Claude / GLM 5.2 closed-book. Paired with

Opal's memory + tools + test-time compute it punches far above its size on the axes an on-device system

owns — privacy, cost, persistence, uncensored, long-context efficiency, and your private corpus. No

fabricated benchmarks. Details: cognis-digital/cognis-opal.

Built by Cognis Digital · powered by Aleph memory.