ayoubkirouane/Sophea-Titan-1

Sophea-Titan-1Greek fine-tuned multimodal Qwen3.6-27B — non-thinking chat

Sophea-Titan-1 is a 27B Greek fine-tuned multimodal chat model built on Qwen3.6-27B, with register control and a fixed assistant identity — while retaining English ability and vision.

• Creator: Kiefer SA
• Base model: Qwen3.6-27B (multimodal, 64 layers, hidden 5120)
• Languages: Greek (primary), English (retained)
• Decoding: non-thinking (enable_thinking=false) — see recommended sampling under Usage

Serve non-thinking. Set enable_thinking=false in the chat template. Over an OpenAI-compatible vLLM endpoint, pass extra_body={"chat_template_kwargs": {"enable_thinking": false}}. With thinking left on, Greek output quality degrades sharply.

Intended use

• General-purpose Greek conversational assistant / chat, with formal ↔ informal register control
• Greek-knowledge QA; English retained as a secondary language

Not evaluated for safety-critical or legal/medical decisions.

Fine-tuned from Qwen3.6-27B.

---

Evaluation

Non-thinking, greedy (temperature 0). Sophea-Titan-1 in bold.

Models compared:

n/a = not applicable (text-only) · n.t. = not identity-tuned.

Headline scorecard

At a glance

Capability profile — General Greek benchmarks, English and vision across open-weight models (Sophea-Titan-1 in bold; frontier MoEs GLM-5.2 / MiniMax-M3 shown for reference, no vision):

General Greek benchmarks vs. model size — Sophea-Titan-1 has the strongest general-Greek score of the ~27B open models, approaching far larger frontier MoEs at a fraction of the size:

Per-benchmark detail

Three benchmark families in one table — General Greek benchmarks (9), English retention (5), and Vision — MMStar (1500) — on the same five models. n/a = not applicable (text-only, no vision).

greekmmlu — per-subject (31)

Frontier & other API models (reference)

Evaluated via API (thinking disabled where supported): GPT-5.5, Claude Opus 4.8, Gemini 3.5, GLM-5.2, MiniMax-M3. API-model Greek/English benchmarks use letter-answer scoring (local models use log-likelihood). Kimi-K3 added 2026-07-29: kimi-k3 via Moonshot API (temperature=1 forced, thinking cannot be disabled). Inkling-Small added 2026-07-31: thinkingmachines/Inkling-Small-NVFP4 (276B/12B-active open-weight MoE, Apache-2.0), self-hosted via vLLM; thinking on, temperature=1 (same protocol caveat as noted for reasoning-locked APIs). Qwen3.8-Max added 2026-08-04: qwen3.8-max via Qwen Cloud API (thinking disabled, temperature=0 — matches the card protocol). DeepSeek-V4.1-Flash added 2026-09-14: deepseek-ai/DeepSeek-V4.1-Flash (552B/16B-active open-weight MoE + 196B Engram tables, MIT), self-hosted via vLLM on 3× B200 (FP4 experts / FP8 dense as shipped); thinking disabled, temperature=0 — matches the card protocol. 4 of 7,000 answers were unparsed (the model reasoned in-content instead of answering with a letter) and are scored as wrong. Vision — MMStar (1500, same letter-answer protocol, thinking disabled): 0.721 overall (coarse perception 0.756 · fine-grained perception 0.620 · instance reasoning 0.768 · logical reasoning 0.764 · math 0.800 · science & technology 0.620; 11 unparsed scored as wrong) vs Sophea-Titan-1 0.714.

General Greek benchmarks (9)

English — 5 benchmarks

greekmmlu

Quantized variants & other formats

Smaller-footprint and alternate-runtime builds of this model:

• GPU / vLLM (text-only — vision tower not included): Sophea-Titan-1-FP8 — lossless, ≈29 GB · Sophea-Titan-1-NVFP4 — 4-bit, ≈19 GB
• llama.cpp — Sophea-Titan-1-GGUF
• Apple Silicon — Sophea-Titan-1-mlx

The trade-off plot below covers the vLLM builds scored on the Greek suite; GGUF / MLX are format conversions.

VRAM vs. Greek benchmark score across the Sophea-Titan-1 quantized family (bf16 → FP8 → NVFP4).

Usage

Serve with vLLM (OpenAI-compatible; 27B bf16 ≈ 54 GB — add --tensor-parallel-size N for multi-GPU):

vllm serve KIEFERSA/Sophea-Titan-1 --served-model-name sophea-titan-1 --trust-remote-code \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3

For text-only serving (skips the vision tower — much faster), add --language-model-only.

Serve with vision enabled — raise the per-request image budget explicitly:

vllm serve KIEFERSA/Sophea-Titan-1 --served-model-name sophea-titan-1 --trust-remote-code \
  --limit-mm-per-prompt '{"image": 4}' \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3

resp = client.chat.completions.create(
    model="sophea-titan-1",
    messages=[{"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
        {"type": "text", "text": "Περίγραψε την εικόνα στα ελληνικά."},
    ]}],
    temperature=0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},   # required: non-thinking
)
print(resp.choices[0].message.content)

Recommended sampling (instruct / non-thinking): temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0.

Client (OpenAI SDK) — remember non-thinking:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="sophea-titan-1",
    messages=[{"role": "user", "content": "Ποια είναι η πρωτεύουσα της Ελλάδας;"}],
    temperature=0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},   # required: non-thinking
)
print(resp.choices[0].message.content)

Transformers (multimodal — load with the image-text-to-text head to keep vision):

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

proc = AutoProcessor.from_pretrained("KIEFERSA/Sophea-Titan-1", trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    "KIEFERSA/Sophea-Titan-1", torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)

messages = [{"role": "user", "content": [{"type": "text",
             "text": "Ποια είναι η πρωτεύουσα της Ελλάδας;"}]}]
text = proc.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = proc(text=[text], return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(proc.batch_decode(out[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])

Speculative decoding (MTP)

This model ships the multi-token-prediction head — 15 mtp.* tensors (~0.85 GB, bf16) in model-mtp.safetensors. This is the single-layer draft stack that config.json has always declared through text_config.mtp_num_hidden_layers: 1.

Earlier revisions did not contain it. The LoRA merge loaded the base through AutoModelForCausalLM, and transformers declares _keys_to_ignore_on_load_unexpected = [r"^mtp.*"] for this architecture, so the head was discarded at load and never written back — the config advertised a module the weights did not contain. The head has been restored tensor-for-tensor from the base, leaving every other weight untouched; the addition is purely additive, so existing serve commands keep working unchanged. To pin the previous bytes, use revision="726f410a802dcf7a720dca898da9c7b91943911e".

Enable it with vLLM (≥ 0.23.0):

vllm serve KIEFERSA/Sophea-Titan-1 --served-model-name sophea-titan-1 --trust-remote-code \
  --speculative-config '{"method":"qwen3_5_mtp","num_speculative_tokens":1}'

Provenance. These are the base Qwen3.6-27B MTP weights. The Greek SFT and DPO LoRA never targeted mtp.*, so no fine-tuned draft head exists. This cannot affect output quality: speculative decoding verifies every drafted token against the main model, so a stale drafter changes throughput only, never the output distribution.

The -FP8 and -NVFP4 variants ship the same head and support the same flag; on those builds the MTP Linears are held in bf16 and listed in quantization_config.ignore.

License

Inherits the Qwen3.6-27B base-model license. Verify base-model terms before use.