KIEFERSA/Sophea-Genesis-1

Sophea-Genesis-1

Language model built to generate well-formed Modern Greek,
while retaining general Greek and English capabilities

Sophea-Genesis-1 (from Greek γένεσις, "origin, creation"; for a model built to generate well-formed Modern Greek) is an open-weight 27B-parameter Greek and English language model.

• Creator: Kiefer SA
• Base model: Qwen3.6-27B
• Languages: Greek, English
• License: Apache-2.0

Highlights

• Leading Greek morphology among comparably-sized open models. The most accurate of the open models we evaluated at producing correctly inflected Greek word-forms.
• Broad competence intact. Matches or beats similarly sized open models on Greek and English knowledge, reading comprehension, and instruction following.
• Open weights, runs on-prem. Apache-2.0; no text leaves your servers, no per-request fee.

Our Greek-morphology result is measured on an open benchmark, released alongside the paper that introduces it:

Benchmark

MORFES

500 expert-verified items of Modern Greek inflectional morphology.

📄 Read the paper (arXiv) 📊 Browse the dataset

Overview

Accurate Greek word-formation. Modern Greek is a highly inflected language: every noun, adjective, and verb takes many distinct forms, varying by person, number, gender, case, tense, aspect, mood, and voice. Producing the correct form is a requirement of well-formed Greek, and our evaluation shows that general-purpose models fall short of it: models that identify the correct form among candidates (recognition) fail considerably more often when asked to write the form itself (generation). On MORFES, the 500-item expert-verified benchmark released alongside the model, Sophea-Genesis-1 attains the highest accuracy in both recognition and generation among the open models we evaluated, including comparably sized general models (Qwen3-32B, Gemma-4-31B-it, EuroLLM-22B), the open Greek models Llama-Krikri-8B and Meltemi-7B, and our own general-purpose model Sophea-Titan-1.

General capability retained. The gains on Greek inflection do not come at the cost of broad competence. On general-capability benchmarks of knowledge, reading comprehension, and instruction following, in both Greek and English, Sophea-Genesis-1 is comparable to or ahead of similarly sized open models, with all models evaluated under identical settings; the results below report the full comparison.

Use cases

Sophea-Genesis-1 is released as open weights under the Apache-2.0 license, which permits use, modification, and commercial deployment; the terms of that license govern all use of the model. Because the weights are open, the model runs entirely on an organization's own infrastructure: no text is transmitted to a third party, and no per-request fee is incurred. For many Greek-language applications this is a deciding property, since organizations in government, law, finance, healthcare, education, and publishing often face constraints on the location where their data is processed.

Sophea-Genesis-1 is a general-purpose model whose general capability in Greek and English is comparable to that of similarly sized open models, and whose distinguishing property is the accuracy of its Greek morphological inflection, the production of correctly formed words. It is therefore suited to applications in which the quality of Greek output is a deciding requirement, particularly where that text cannot be sent to an external service. Such applications include, among others:

• Greek-language assistants: user-facing or internal assistants for organizations that operate in Greek and keep their data in-house, including the drafting of correspondence, reports, and administrative text under human supervision.
• Educational and language-learning tools: applications that present learners with correct Greek word-forms, including the complete inflectional paradigm of a given noun, adjective, or verb.

It is also a base to build on: because MORFES is released with the model, a team adapting it to a specific domain through further post-training can re-evaluate the adapted model on the benchmark and verify that its accuracy on Greek inflection is preserved.

Results

Morphology (MORFES)

Model	Recognition	Production (per form)	Production (per item)
Sophea-Genesis-1 (27B)	0.912	0.900	0.840
Qwen3-32B	0.774	0.523	0.350
Gemma-4-31B-it	0.640	0.685	0.526
Sophea-Titan-1 (27B)	0.842	0.798	0.714
Qwen3.6-27B	0.818	0.720	0.630
EuroLLM-22B-Instruct	0.836	0.385	0.360
Llama-Krikri-8B-Instruct	0.816	0.305	0.302
Meltemi-7B-Instruct	0.768	0.094	0.136

MORFES comprises 500 items, each posed in two modes. In recognition the four candidate forms are shown and the model selects one; in production the same item is posed with no candidates and the model writes the form(s) itself. Both modes are 0-shot. We report three metrics:

• Recognition (acc_norm) — each candidate form is scored as a continuation of the question by its log-likelihood, normalized for length so forms of unequal length compete fairly; we report the proportion of items whose highest-scoring candidate is the correct one.
• Production (per form) — the proportion of target forms produced correctly, counting a form correct when it exactly matches one of the accepted answers for that cell, accentuation included (some cells admit more than one attested form).
• Production (per item) — the same generations scored all-or-nothing: an item counts only if every one of its forms is correct, placing production on the same per-item basis as recognition.

General-capability panel

Model	Size	Knowledge (MMLU, EN)	Knowledge (GreekMMLU, EL)	Reading (Belebele, EN)	Reading (Belebele, EL)	Instruction (IFEval, EN)
Sophea-Genesis-1 (27B)	27B	0.866	0.834	0.963	0.933	0.837
Qwen3-32B	33B	0.819	0.651	0.946	0.872	0.834
Gemma-4-31B-it	31B	0.633	0.818	0.960	0.923	0.909
Sophea-Titan-1 (27B)	27B	0.864	0.834	0.958	0.928	0.774
Qwen3.6-27B	27B	0.866	0.795	0.958	0.918	0.835
EuroLLM-22B-Instruct	22B	0.681	0.724	0.820	0.784	0.712
Llama-Krikri-8B-Instruct	8B	0.637	0.673	0.877	0.850	0.811
Meltemi-7B-Instruct	7B	0.557	0.637	0.651	0.576	0.351

All reported scores are our own measurement under one fixed setup (lm-evaluation-harness v0.4.12 on a vLLM backend, greedy and non-thinking; multiple-choice cells scored by log-likelihood, generative cells with the model's chat template), not an effort to reproduce the figures each model's authors report under their own harness, prompt format, or metric.

Evaluation details.

• Knowledge (MMLU, EN) — MMLU (Hendrycks et al., 2021), the standard English knowledge exam: multiple-choice questions across 57 academic and professional subjects. Measured 5-shot under the canonical protocol: each test question is preceded by five worked examples (a question with its correct answer) taken from that subject's own dedicated dev split, which holds exactly five examples per subject, and never from another subject. Scored by accuracy.
• Knowledge (GreekMMLU, EL) — GreekMMLU (Zhang et al., 2026; dascim/GreekMMLU), the Greek knowledge exam: multiple-choice questions natively authored in Greek (not machine-translated from English). Measured 0-shot (no in-context examples); accuracy.
• Reading (Belebele, EN) — Belebele (Bandarkar et al., 2024; eng_Latn), reading comprehension in English: a short passage followed by a four-option question about it. Measured 0-shot (no in-context examples); accuracy.
• Reading (Belebele, EL) — Belebele (ell_Grek), the same reading-comprehension task in Greek. Measured 0-shot (no in-context examples); accuracy.
• Instruction (IFEval, EN) — IFEval (Zhou et al., 2023), instruction-following: whether a response obeys verifiable format constraints (for example a word limit, or a required keyword). Measured 0-shot. Scored by prompt-level strict accuracy: the share of prompts for which every instruction is satisfied (prompt-level, so a prompt counts only if all of its instructions pass), judged on the output verbatim.

Usage

Serve as a text model with vLLM. The model inherits a vision pathway from its base architecture, which this release neither uses nor evaluates, so serve with --language-model-only:

vllm serve KIEFERSA/Sophea-Genesis-1 --served-model-name sophea-genesis-1 --trust-remote-code --language-model-only --mamba-cache-dtype float16

Query the endpoint with any OpenAI-compatible client; non-thinking mode is selected through the chat template:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="sophea-genesis-1",
    messages=[{"role": "user", "content": "Δώσε την κλίση του ρήματος «σχηματίζω» στον αόριστο."}],
    temperature=0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)

The results above were measured with greedy decoding (temperature=0) in non-thinking mode. For open-ended generation, the base model recommends the following non-thinking sampling parameters: temperature=0.7, top_p=0.8, top_k=20, presence_penalty=1.5.

Limitations

Like all large language models, Sophea-Genesis-1 may generate incorrect or outdated factual statements. In high-stakes settings, its outputs should be subject to human review rather than relied upon as an authoritative source of fact.

License

Apache-2.0 (see LICENSE).