GGUF (llama.cpp) builds of the text language model of KIEFERSA/Sophea-Titan-1 (Qwen3.6-27B, hybrid GatedDeltaNet + full attention). Vision is not carried into GGUF. Non-thinking, greedy. Requires a llama.cpp build with qwen35 support.

Files

Sophea-Titan-1.Q4_K_M.gguf — 16.5 GB
Sophea-Titan-1.Q5_K_M.gguf — 19.2 GB
Sophea-Titan-1.Q8_0.gguf — 28.6 GB

Runtime compatibility

These are the text language model only (no vision). The architecture is Qwen3.6/Qwen3.5's hybrid GatedDeltaNet + full attention, so the runtime's bundled llama.cpp must be recent enough to support the qwen35 architecture. If a runtime errors with "unknown model architecture: qwen35", it simply needs a newer llama.cpp.

Runtime	Status
llama.cpp (recent build)	✅ Verified — loads and runs
LM Studio	✅ Works when its runtime is a `qwen35`-capable `llama.cpp` — use the latest LM Studio and select the newest runtime
Ollama (≥ 0.32)	✅ Works — import the GGUF; serve chat with `think: false` (non-thinking model)

System requirements (local / CPU)

Runs on a normal machine — no GPU required. You need roughly the file size in RAM, plus a couple GB.

Quant	Size	Min RAM	Comfortable
Q4_K_M	16.5 GB	24 GB	32 GB
Q5_K_M	19.2 GB	24 GB	32 GB
Q8_0	28.6 GB	32 GB	48 GB

CPU: any modern x86-64 (AVX2) or Apple Silicon; 8+ cores for usable speed. A GPU is optional (faster), not required.

Usage

Download a quant (private repo — use your HF token):

bash

huggingface-cli download KIEFERSA/Sophea-Titan-1-GGUF Sophea-Titan-1.Q4_K_M.gguf \
  --local-dir ./sophea --token $HF_TOKEN

llama.cpp (verified) — non-thinking, greedy:

bash

llama-cli -m ./sophea/Sophea-Titan-1.Q4_K_M.gguf \
  -p "Ποια είναι η πρωτεύουσα της Ελλάδας;" -n 64 --temp 0

LM Studio — place under publisher/model/, then pick it in the model loader:

bash

mkdir -p ~/.lmstudio/models/KIEFERSA/Sophea-Titan-1-GGUF
cp ./sophea/Sophea-Titan-1.Q4_K_M.gguf ~/.lmstudio/models/KIEFERSA/Sophea-Titan-1-GGUF/

Ollama (≥ 0.32 — older versions error with "unknown architecture: qwen35"):

bash

printf 'FROM ./sophea/Sophea-Titan-1.Q4_K_M.gguf\nPARAMETER temperature 0\n' > Modelfile
ollama create sophea-titan-1 -f Modelfile
ollama run sophea-titan-1 --think=false "Ποια είναι η πρωτεύουσα της Ελλάδας;"

Non-thinking model: pass think: false (CLI --think=false, or "think": false in the API) — otherwise the chat reply comes back empty. Ollama also serves an OpenAI-compatible API at http://localhost:11434/v1.

Serve an OpenAI-compatible API with llama.cpp's llama-server:

bash

llama-server -m ./sophea/Sophea-Titan-1.Q4_K_M.gguf \
  --host 0.0.0.0 --port 8080 --api-key sk-local -c 8192

Call it with the OpenAI SDK — just point base_url at the server:

python

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="sk-local")
resp = client.chat.completions.create(
    model="sophea-titan-1",            # any name — the loaded GGUF is served
    messages=[{"role": "user", "content": "Ποια είναι η πρωτεύουσα της Ελλάδας;"}],
    temperature=0,
)
print(resp.choices[0].message.content)

Or with plain curl:

bash

curl http://localhost:8080/v1/chat/completions \
  -H "Authorization: Bearer sk-local" -H "Content-Type: application/json" \
  -d '{"model":"sophea-titan-1","messages":[{"role":"user","content":"Γεια σου"}],"temperature":0}'

Quants: Q4_K_M (16.5 GB, recommended) · Q5_K_M (19.2 GB) · Q8_0 (28.6 GB, near-lossless).

License

Inherits the Qwen3.6-27B base-model license.