Muninn

Muninn is a 346M-parameter multilingual retriever for natural-language-query → code-function retrieval, with an 8,192-token serving context and 2,048-dimensional embeddings.

It is trained from voyageai/voyage-4-nano and uses a bidirectional Qwen3 encoder. Muninn supports Matryoshka truncation at 512, 1,024, 1,536, and 2,048 dimensions. We recommend 2,048 dimensions: that is the native evaluation setting, and truncating to 512 cost roughly two recall points at depth on Quarry.

Usage

Muninn requires trust_remote_code=True because this repository includes the custom Qwen3BidirectionalModel implementation.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "BrokkAI/Muninn",
    trust_remote_code=True,
    truncate_dim=2048,
)
model.max_seq_length = 8192

queries = ["Where is retry backoff calculated for failed HTTP requests?"]
documents = [
    "src/net/client.py/HttpClient/retry_delay\n"
    "class HttpClient:def retry_delay(self, attempt):\n"
    "    return min(60, 2 ** attempt)"
]

query_embeddings = model.encode(queries, prompt_name="query", normalize_embeddings=True)
document_embeddings = model.encode(
    documents, prompt_name="document", normalize_embeddings=True
)
scores = model.similarity(query_embeddings, document_embeddings)
print(scores)

The exact inference prompts are:

query:    Represent the query for retrieving supporting documents: 
document: Represent the document for retrieval: 

Document format

Quarry results use the header format produced by swerank_document_text() before the document prompt is applied. For a free function:

{path}/{function_name}
{source}

For a class method:

{path}/{ClassName}/{function_name}
class {ClassName}:{source}

The worked usage example above is therefore the exact representation of a method named HttpClient.retry_delay in src/net/client.py. Preserve this shape when comparing against reported results.

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. Muninn leads a field that includes much larger open models and commercial APIs.

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:

LocBench — 560 Python issues, function-level accuracy:

CoIR text-to-code — NDCG@10:

CodeSearchNet — mean NDCG@10 over six languages (Python, JavaScript, Go, Ruby, Java, PHP; 1,000 queries per language):

Languages

C, C++, C#, Go, Java, JavaScript, PHP, Python, Rust, Scala, and TypeScript.

License and attribution

Muninn is released under the Apache License 2.0. It is derived from voyageai/voyage-4-nano, also released under Apache-2.0. See LICENSE for the full license text.