ogma-small

Efficient English text embedding model for semantic search, RAG, vector search, retrieval, clustering, classification, STS, and agent memory — MTEB 56.32, 8.6M parameters, 1024-token context

Maintained successor (MIT): the permissively licensed model in this family — including the CoreML build that used to sit under coreml/ here — is axiotic/open-ogma-small. This repository's own weights remain CC-BY-NC-4.0 and it no longer hosts that artifact.

Ogma Small is the flagship efficiency model in the family. At 8.6M parameters it scores 56.32 MTEB in our canonical 66-task Ogma paper results, while using only 38% of MiniLM-L6-v2's parameters, running 1.75× faster on CPU, and handling inputs 4× longer (1024 vs 256 tokens). Purpose-built to be the drop-in for every place you currently reach for MiniLM.

Why the name Ogma?

Ogma is named after Ogma (also written Oghma), the Irish god associated with eloquence and credited in myth with inventing Ogham, an early alphabet for encoding language into symbols. That is the core job of an embedding model: turn language into compact vectors that machines can search, compare, cluster, and reason over.

Use cases

ogma-small is the default efficiency model for semantic search, RAG retrieval, agent memory, vector databases, document retrieval, text classification, clustering, STS / sentence similarity, and lightweight reranking pipelines. It is aimed at teams looking for a small, fast MiniLM-style embedding model with longer context and strong MTEB quality.

Good fits:

• On-device or local-first applications where MiniLM-class quality is useful but model size, CPU latency, and privacy matter.
• Production RAG systems that need affordable embeddings for documents, chunks, tickets, chats, and internal knowledge bases.
• Agent memory and tool-use systems where frequent embedding calls should stay cheap and local when possible.
• Vector search at scale where smaller models and Matryoshka sub-dimensions can reduce index size and query cost.
• Classification and clustering features for safety filters, routing, topic grouping, deduplication, and analytics.

Choose ogma-small when you want the best balance of quality, speed, size, and deployability across edge, local, and server workloads.

Highlights

• 🏆 MTEB avg 56.32 — canonical Ogma paper result over 66/66 MTEB English tasks
• ⚡ 1.75× faster than MiniLM on single-threaded CPU inference (92.9 vs 53.1 docs/s)
• 📏 1024-token context — 4× longer than all-MiniLM-L6-v2 (256 tokens)
• 🔀 Symmetric routing via task tokens — encode everything with [SYM], or use [QRY]/[QRY] for retrieval (queries and documents both encoded with task="qry"); benchmark both routes on your task
• 📐 Matryoshka dims: [256, 128, 64, 32] — one model, any precision
• 🛡️ +4.0% F1 on prompt injection detection vs MiniLM (same architecture series)

Performance

MTEB English — 66/66 tasks (category-averaged)

Benchmarked with MTEB v2.10.7 on the standard 66-task English benchmark using category averaging (same methodology as the MTEB leaderboard).

Why choose Ogma Small?

ogma-small is the default recommendation for most use cases. It is MiniLM-class quality while being faster, smaller, and context-aware. Use ogma-base when you need the extra quality margin; use ogma-mini when you need to go sub-4M parameters.

Safety — Toxicity & Prompt Injection Detection

Evaluated on the Ogma transformer architecture (same family). Embeddings are extracted then fed to a logistic regression (LR) or MLP classifier head — the embedding model itself is not fine-tuned. Evaluated against all-MiniLM-L6-v2 as baseline.

1. Jigsaw Toxic Comment Classification

Dataset: Arsive/toxicity_classification_jigsaw — Binary toxicity classification Train: 25,960 · Test: 6,490

Ogma (LR) leads MiniLM (LR) by +2.01% F1. MiniLM (MLP) leads on this dataset — the additional training data (25K samples) allows the MLP to compensate for MiniLM's slightly weaker base representations.

2. Prompt Injection Detection — deepset/prompt-injections

Dataset: deepset/prompt-injections — Binary injection detection Train: 546 · Test: 116 (low-data regime)

Ogma leads across both classifiers: +4.03% F1 (MLP), +4.23% F1 (LogReg). Ogma's representations are better separated in the low-data regime — it achieves 100% precision with LogReg, meaning zero false positives.

3. Prompt Injection Detection — neuralchemy/Prompt-injection-dataset

Dataset: neuralchemy/Prompt-injection-dataset — Binary injection detection Train: 4,391 · *Test: