Muninn-small
Muninn-small is a 47M-parameter multilingual retriever for natural-language-query →
code-function retrieval, with an 8,192-token context and 384-dimensional embeddings.
It is the efficient Muninn tier, trained from
ibm-granite/granite-embedding-small-english-r2 with a ModernBERT encoder.
Usage
1from sentence_transformers import SentenceTransformer
2
3model = SentenceTransformer("BrokkAI/Muninn-small")
4model.max_seq_length = 8192
5
6queries = ["Where is retry backoff calculated for failed HTTP requests?"]
7documents = [
8 "src/net/client.py/HttpClient/retry_delay\n"
9 "class HttpClient:def retry_delay(self, attempt):\n"
10 " return min(60, 2 ** attempt)"
11]
12
13query_embeddings = model.encode(queries, prompt_name="query", normalize_embeddings=True)
14document_embeddings = model.encode(
15 documents, prompt_name="document", normalize_embeddings=True
16)
17scores = model.similarity(query_embeddings, document_embeddings)
18print(scores)
The exact prompts used during training and evaluation are:
1query: Given a GitHub issue, retrieve code that must be changed to fix it.
2 Query:
3document: Passage: Code chunk from repository.
The query prompt literally says “GitHub issue” because that is the trained prefix. Quarry
queries are synthetic natural-language behavioral descriptions, not raw issue text. Keep
the prefix unchanged at inference time even when the input is another kind of natural-
language code-search query.
Document format
Quarry results use the header format produced by swerank_document_text() before the
document prompt is applied. For a free function:
1{path}/{function_name}
2{source}
For a class method:
1{path}/{ClassName}/{function_name}
2class {ClassName}:{source}
The worked usage example above is therefore the exact representation of a method named
HttpClient.retry_delay in src/net/client.py.
Quarry results
Quarry contains 6,525 synthetic behavioral
queries over real repository revisions. The metric is strict all-gold micro recall@k: for
each query,
|gold ∩ top-k| / |gold|, followed by a flat mean over queries. Models use their
native dimensions and the header document format above.
We measured every row ourselves with the Quarry harness. At 47M parameters, Muninn-small
beats its own base by 4.6 points at recall@5 and matches SweRank-Small (137M, Python-only
training) at recall@20 while running at roughly a third of its size.
| Model | Params | recall@5 | recall@20 | recall@50 |
|---|
| Muninn | 346M | 61.4 | 82.6 | 90.7 |
| voyage-code-3 | API | 59.1 | 81.8 | 90.4 |
| voyage-4 | API | 59.0 | 80.8 | 90.0 |
| SweRank-Large | 7B | 58.5 | 80.6 | 89.6 |
| Nemotron-3-Embed-1B | 1B | 58.3 | 79.9 | 88.7 |
| bge-code-v1 | 1.5B | 58.2 | 79.8 | 88.7 |
| voyage-4-nano (Muninn base) | 346M | 57.5 | 79.7 | 89.0 |
| voyage-4-lite | API | 57.3 | 79.6 | 89.0 |
| Qwen3-Embedding-8B | 8B | 56.0 | 78.2 | 88.1 |
| SweRank-Small | 137M | 53.5 | 74.6 | 83.9 |
| Muninn-small | 47M | 52.3 | 74.6 | 84.7 |
| text-embedding-3-large | API | 51.7 | 74.3 | 85.0 |
| Qwen3-Embedding-0.6B | 0.6B | 50.9 | 72.8 | 83.8 |
| granite-embedding-small-english-r2 (base) | 47M | 47.7 | 69.7 | 80.5 |
Other benchmarks
† marks published numbers (SweRank paper for the localization benchmarks; the CoIR
leaderboard and the Qwen3-Embedding paper elsewhere). Unmarked rows are our measurements
on the same protocol as the corresponding published numbers.
SWE-Bench-Lite localization — 274 Python issues, function-level accuracy:
| Model | Acc@5 | Acc@10 |
|---|
| Muninn (346M) | 73.4 | 79.6 |
| SweRank-Large (7B)† | 71.9 | 82.1 |
| SweRank-Small (137M)† | 63.1 | 74.5 |
| Muninn-small (47M) | 48.5 | 61.0 |
LocBench — 560 Python issues, function-level accuracy:
| Model | Acc@10 | Acc@15 |
|---|
| SweRank-Large (7B)† | 63.2 | 67.3 |
| Muninn (346M) | 60.2 | 65.7 |
| SweRank-Small (137M)† | 58.6 | 63.4 |
| Muninn-small (47M) | 46.4 | 51.6 |
CoIR text-to-code — NDCG@10:
| Model | APPS | CosQA |
|---|
| bge-code-v1 (1.5B)† | 98.1 | 46.7 |
| voyage-code-3 (API)† | 93.6 | 34.5 |
| Qwen3-Embedding-8B† | 91.1 | 38.0 |
| Muninn (346M) | 77.4 | 31.3 |
| Qwen3-Embedding-0.6B† | 75.3 | 36.5 |
| Muninn-small (47M) | 12.4 | 35.1 |
CodeSearchNet — mean NDCG@10 over six languages (Python, JavaScript, Go, Ruby, Java,
PHP; 1,000 queries per language):
| Model | Mean NDCG@10 |
|---|
| Qwen3-Embedding-8B† | 92.7 |
| Qwen3-Embedding-0.6B† | 91.0 |
| Muninn (346M) | 90.1 |
| SweRank-Small (137M) | 87.9 |
| SweRank-Large (7B) | 85.0 |
| Muninn-small (47M) | 77.8 |
Muninn-small is specialized for repository-scale function retrieval; the APPS number shows
it is not a general text-to-code model for competitive-programming-style queries.
Languages
C, C++, C#, Go, Java, JavaScript, PHP, Python, Rust, Scala, and TypeScript.
License and attribution
Muninn-small is released under the Apache License 2.0. It is derived from
ibm-granite/granite-embedding-small-english-r2,
also released under Apache-2.0. See
LICENSE for the full license text.
