KIEFERSA/Sophea-Titan-1

Sophea-Titan-1

Greek 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:

column	model	what it is	identity-tuned?
Base	Qwen3.6-27B	the base Sophea-Titan-1 is built on	no
Sophea-Titan-1	this model	Qwen3.6-27B + general-purpose Greek conversational SFT	yes
Sophea-K1	KIEFERSA/sophea-k1	KIEFERSA production multimodal Greek model (sophea.ai)	yes
Qwen3-30B-A3B	Qwen/Qwen3-30B-A3B	text-only MoE (30B/3B-active), not fine-tuned	no
Qwen3.6-35B-A3B	Qwen/Qwen3.6-35B-A3B	multimodal MoE (35B / 3B-active), not fine-tuned	no
Krikri-8B	ilsp/Llama-Krikri-8B-Instruct	Llama-3.1-8B Greek LLM (external baseline)	no

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

Headline scorecard

Axis	Base	Sophea-Titan-1	Sophea-K1	Qwen3-30B-A3B	Qwen3.6-35B-A3B	Krikri-8B
General Greek benchmarks — macro (9)	0.7157	0.7369	0.7240	0.6346	0.6983	0.5977
English retention — macro (5)	0.8631	0.8783	0.8702	0.7631	0.8362	0.7423
Vision — MMStar (1500)	0.6533	0.7140	0.6620	n/a	0.590	n/a

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).

Benchmark	Base	Sophea-Titan-1	Sophea-K1	Qwen3-30B-A3B	Qwen3.6-35B-A3B	Krikri-8B
General Greek benchmarks (9)
greekmmlu	0.849	0.854	0.855	0.753	0.838	0.675
mmlu_greek	0.798	0.797	0.782	0.628	0.770	0.519
hellaswag	0.611	0.680	0.622	0.410	0.574	0.572
medical_mcqa	0.312	0.387	0.368	0.232	0.312	0.287
winogrande	0.590	0.626	0.599	0.559	0.564	0.609
arc_challenge	0.944	0.950	0.945	0.867	0.916	0.690
arc_easy	0.973	0.973	0.971	0.931	0.969	0.832
belebele	0.941	0.950	0.938	0.887	0.926	0.771
truthfulqa	0.423	0.415	0.437	0.446	0.416	0.425
MACRO (9)	0.7157	0.7369	0.7240	0.6346	0.6983	0.5977
English retention (5)
arc_challenge	0.972	0.979	0.977	0.932	0.959	0.765
arc_easy	0.990	0.992	0.994	0.981	0.990	0.892
hellaswag	0.764	0.805	0.780	0.500	0.734	0.751
mmlu	0.858	0.858	0.858	0.745	0.840	0.608
winogrande	0.732	0.758	0.742	0.658	0.658	0.695
MACRO (5)	0.8631	0.8783	0.8702	0.7631	0.8362	0.7423
Vision — MMStar (1500)
coarse perception	0.744	0.728	0.728	n/a	0.704	n/a
fine-grained perception	0.636	0.608	0.628	n/a	0.588	n/a
instance reasoning	0.756	0.792	0.792	n/a	0.740	n/a
logical reasoning	0.668	0.764	0.636	n/a	0.572	n/a
math	0.464	0.708	0.500	n/a	0.372	n/a
science & technology	0.652	0.684	0.688	n/a	0.564	n/a
OVERALL	0.653	0.714	0.662	n/a	0.590	n/a

greekmmlu — per-subject (31)

subject	Base	Sophea-Titan-1	Sophea-K1	Qwen3-30B-A3B	Qwen3.6-35B-A3B	Krikri-8B
Accounting	0.864	0.864	0.859	0.761	0.848	0.663
Agriculture	0.853	0.870	0.853	0.721	0.819	0.707
Art	0.748	0.756	0.770	0.621	0.757	0.625
Biology	0.864	0.868	0.859	0.794	0.856	0.672
Chemistry	0.827	0.753	0.778	0.654	0.790	0.519
Civil Engineering	0.793	0.827	0.819	0.678	0.785	0.637
Clinical Knowledge	0.810	0.820	0.795	0.689	0.790	0.686
Computer Networks & Security	0.730	0.762	0.714	0.540	0.667	0.476
Computer Science	0.883	0.880	0.866	0.827	0.869	0.768
Driving Rules	0.837	0.820	0.829	0.756	0.811	0.631
Economics	0.911	0.926	0.926	0.840	0.894	0.670
Education	0.857	0.867	0.895	0.762	0.901	0.687
Electrical Engineering	0.830	0.822	0.844	0.681	0.817	0.538
General Knowledge	0.769	0.792	0.823	0.701	0.781	0.630
Geography	0.943	0.964	0.955	0.870	0.950	0.872
Government and Politics	0.944	0.941	0.961	0.901	0.944	0.899
Greek History	0.871	0.877	0.878	0.711	0.884	0.816
Greek Literature	0.714	0.714	0.643	0.429	0.500	0.500
Greek Mythology	0.870	0.857	0.845	0.744	0.849	0.697
Greek Traditions	0.880	0.891	0.880	0.755	0.867	0.710
Law	0.718	0.703	0.697	0.580	0.677	0.504
Management	0.824	0.841	0.842	0.719	0.814	0.694
Maritime Safety & Rescue	0.703	0.669	0.662	0.608	0.682	0.507
Mathematics	0.897	0.924	0.904	0.844	0.857	0.516
Medicine	0.893	0.895	0.886	0.757	0.882	0.663
Modern Greek Language	0.916	0.918	0.928	0.836	0.911	0.780
Physics	0.838	0.851	0.851	0.769	0.838	0.684
Prehistory	1.000	1.000	0.968	0.984	0.984	0.905
World History	0.950	0.950	0.900	0.850	0.950	0.800
World Religions	0.768	0.755	0.787	0.652	0.800	0.684
OVERALL	0.849	0.854	0.855	0.753	0.838	0.675

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).

Axis	Sophea-Titan-1	GPT-5.5	Claude-4.8	Gemini-3.5	GLM-5.2	MiniMax-M3	Kimi-K3	Inkling-Small	Qwen3.8-Max
General Greek benchmarks macro (9)	0.737	0.909	0.922	0.872	0.810	0.821	0.916	0.881	0.898
English macro (5)	0.878	0.925	0.940	0.911	0.892	0.890	0.947	0.922	0.935

General Greek benchmarks (9)

benchmark	Sophea-Titan-1	GPT-5.5	Claude-Opus-4.8	Gemini-3.5	GLM-5.2	MiniMax-M3	Kimi-K3	Inkling-Small	Qwen3.8-Max
greekmmlu	0.854	0.905	0.900	0.898	0.805	0.837	0.902	0.872	0.898
mmlu_greek	0.797	0.888	0.884	0.874	0.769	0.773	0.932	0.898	0.862
hellaswag	0.680	0.891	0.909	0.816	0.693	0.741	0.790	0.666	0.794
medical_mcqa	0.387	0.928	0.917	0.912	0.787	0.833	0.944	0.926	0.902
winogrande	0.626	0.781	0.813	0.641	0.705	0.662	0.880	0.850	0.860
arc_challenge	0.950	0.968	0.966	0.952	0.904	0.924	0.972	0.960	0.966
arc_easy	0.973	0.984	0.982	0.973	0.949	0.960	0.984	0.972	0.974
belebele	0.950	0.953	0.947	0.934	0.913	0.911	0.956	0.954	0.942
truthfulqa	0.415	0.885	0.977	0.845	0.766	0.749	0.882	0.832	0.888
MACRO	0.7369	0.9094	0.9216	0.8717	0.8102	0.8212	0.9158	0.8811	0.8984

English — 5 benchmarks

benchmark	Sophea-Titan-1	GPT-5.5	Claude-Opus-4.8	Gemini-3.5	GLM-5.2	MiniMax-M3	Kimi-K3	Inkling-Small	Qwen3.8-Max
arc_challenge	0.979	0.976	0.974	0.958	0.968	0.967	0.978	0.968	0.976
arc_easy	0.992	0.993	0.994	0.979	0.989	0.986	0.990	0.982	0.986
hellaswag	0.805	0.918	0.944	0.913	0.858	0.854	0.876	0.806	0.892
mmlu	0.858	0.908	0.910	0.896	0.845	0.852	0.946	0.934	0.914
winogrande	0.758	0.830	0.877	0.811	0.801	0.793	0.946	0.922	0.906
MACRO	0.8783	0.9249	0.9399	0.9113	0.8922	0.8903	0.9472	0.9224	0.9348

greekmmlu

model	Sophea-Titan-1	GPT-5.5	Claude-Opus-4.8	Gemini-3.5	GLM-5.2	MiniMax-M3	Kimi-K3	Inkling-Small	Qwen3.8-Max
greekmmlu	0.854	0.905	0.900	0.898	0.805	0.837	0.902	0.872	0.898

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.