12 open-weight families18 public releases
View on Hugging FaceLanguage models that understand your operating context.
Build for Greek and multilingual enterprise applications with full control over your data and deployment.
Model Families
Sophea ASR
English and Greek speech-to-text, including mixed Greek–English speech, for meetings, calls, and recordings.
ASR K1 preview · Latest results · Default datasets · Verified View leaderboard
Open-weight models
- (opens in a new tab)
Sophea Titan 1
Greek-first multimodal chat with English support and formal-to-informal register control. Choose a release for your runtime and modality.
- License
- Apache 2.0
- BF16 · multimodal
- MLX · multimodal
- FP8 · text-only
- NVFP4 · text-only
- GGUF · text-only
- (opens in a new tab)
Sophea Genesis 1
A Greek and English language model focused on Modern Greek morphology, including correctly inflected word forms.
- License
- Apache 2.0
- Model card & weights
- (opens in a new tab)
Sophea K1
A Greek-language model with publicly available weights and published Greek-language evaluation results.
- License
- MIT
- Model card & weights
- (opens in a new tab)
Sophea Qwen3.6
Greek and English reasoning with retained vision capabilities. Available as the original fine-tune and a reinforcement-learning refinement.
- License
- Apache 2.0
- v1 · SFT
- v1.1 · RLVR
- (opens in a new tab)
Sophea OSS v1
A Greek and English reasoning model with separate reasoning and answer channels, trained to reason in the language of the question.
- License
- Apache 2.0
- Model card & weights
- (opens in a new tab)
Sophea Nemo 3 Nano v1
Greek and English reasoning built on a hybrid mixture-of-experts architecture, with reasoning matched to the question’s language.
- License
- Apache 2.0 additions
Base-model terms also apply.
- Model card & weights
- (opens in a new tab)
Sophea Nemo 3.5 Lightning v1
Greek and English reasoning with language-matched responses and a multi-token prediction head for speculative decoding.
- License
- Apache 2.0 additions
Base-model terms also apply.
- Model card & weights
- (opens in a new tab)
Sophea Nemo Embedding
Dense Greek and English embeddings for semantic search and retrieval-augmented generation. Represent queries and documents for relevance matching.
- License
- OpenMDW 1.1
- Model card & weights
- (opens in a new tab)
Sophea Nemo Reranker
A Greek and English cross-encoder that reranks retrieved passages by relevance to the query, refining the context passed to a language model.
- License
- NVIDIA Open Model License
Llama 3.2 Community License also applies; see the model card and base-model terms.
- Model card & weights
- (opens in a new tab)
Sophea RAG Nemo3
Document-grounded answer generation in Greek and English, trained to cite sources and abstain when the supplied evidence is insufficient.
- License
- NVIDIA Open Model License
See the model card and base-model terms.
- Model card & weights
- (opens in a new tab)
Sophea Canary ASR
Greek and English speech recognition, plus Greek-to-English speech translation. Standard and MLX weight releases are available.
- License
- CC BY 4.0
- Standard
- MLX
- (opens in a new tab)
Sophea Nano Policy
LIBERO Greek v1: an experimental Greek and English vision-language-action policy for robot manipulation, trained and evaluated in simulation.
- License
- NVIDIA Open Model License
See the model card and base-model terms.
- Model card & weights
Public weights are available under the terms of each release. Model cards describe supported formats, evaluation methods, limitations, and applicable base-model licenses.
Use Case
Answers that work in context.
Support enterprise questions with grounded responses and clear paths to human review.
Explore in a pilotTurn documents into structured work.
Identify requested fields, review uncertain results, and pass approved outputs into a workflow.
Explore in a pilotBring the important details forward.
Create concise summaries of long documents while keeping evidence available for review.
Explore in a pilotConnect intelligence to approved actions.
Define permitted tools, approval boundaries, and escalation before automating a workflow.
Explore in a pilotEvaluation
Agree the criteria.
Evaluate against representative tasks before choosing a model.
- Language & domain
- Greek, English, and your terminology
- Grounding
- Evidence, citations, and abstention
- Task quality
- A rubric agreed with reviewers
- Operating fit
- Latency, cost, and deployment needs
Learn from every review.
Review failure cases, refine the setup, and repeat with the same evaluation criteria.
Enterprise Controls
Your data
Define approved inputs, source access, and handling requirements.
- Isolation by design
- No training on your data
- Data residency options
Access & governance
Scope roles, tools, and review boundaries for the application.
- SSO & SCIM
- RBAC & fine-grained access
- Audit logs & data lineage
Deployment
Align the model runtime with operational requirements.
- On-premises
- Private cloud
- Sovereign cloud
Observability
Identify the evidence needed to inspect quality and behavior.
- Usage & latency
- Quality monitoring
- Drift detection
Developer Quickstart
// Illustrative integration — contract to be confirmed
const answer = await sophea.knowledge.ask({
question: "What are the next steps?",
sources: approvedSources,
identity: currentUser,
includeCitations: true
});
// Review evidence before taking action
renderAnswer(answer.text, answer.citations);# Illustrative request shape; contract to be confirmed
request = {
"question": "What are the next steps?",
"sources": approved_sources,
"identity": current_user,
"include_citations": True
}
# Submit only to your approved endpoint.
# Review evidence before taking action.# Illustrative request shape; not a published API
# Use the endpoint and token supplied for your pilot.
curl "$APPROVED_ENDPOINT" --request POST
# Add your approved authentication header
# and validated request body before sending.// Illustrative request shape; contract to be confirmed
var request = Map.of(
"question", "What are the next steps?",
"sources", approvedSources,
"identity", currentUser,
"includeCitations", true
);
// Use the agreed client and endpoint.
// Review the response and its evidence."citations": ["Pilot brief · p. 2"]