Xunzhuo/vsr-embeddings-v1-text-large is the large member of the vSR text embedding family. It is built for the request path of an AI gateway: predictable task prompts, adjustable embedding width, efficient long-document matching, and a clean interface to parallel safety and routing signals.

Why it is different

Router-native task contract

One semantic model serves retrieval, text matching, clustering, and generic classification embeddings. Query and document roles use explicit task instructions, so applications do not need separate ad-hoc prompt wrappers.

Matryoshka as a latency dial

The same model supports 2560, 1024, 512, 256, 128, 64 dimensions. A deployment can use a smaller vector for hot-path routing and a wider vector for high-accuracy search without changing the model family or index semantics.

Multi-vector long-document retrieval

Long inputs are represented by fixed 320-token chunks with 64-token overlap. Documents are scored with query-to-chunk max-sim, which preserves local evidence that a single pooled vector can dilute. The runtime contract covers documents up to 32K tokens while keeping each model forward short and batchable.

Signal-compatible by design

The embedding model runs in parallel with vSR's domain, safety, fact-check, feedback, modality, and PII extractors. The decision layer keeps those signals separate and inspectable, while this model supplies the shared semantic geometry. That preserves explainability and lets the backend catalog scale independently.

Two deployment tiers

Variant	Parameters	Native dim	MTEB Mean(Task)	MI300X batch-64	Recommended use
small	0.60B	1024	69.01	1042.33	always-on and high-throughput
large	4.02B	2560	69.45	423.11	highest-quality matching

Quality

Evaluation	Result
MTEB Multilingual v2 Mean(Task), 131 tasks	69.45
MTEB Multilingual v2 Mean(TaskType)	60.86
Direct rank in the evaluation snapshot	7
Jina v5 text-small Mean(Task), same snapshot	67.00
Mean(Task) advantage over Jina v5	+2.45
Router multilingual retrieval AUC	0.9770
Router long-document max-sim AUC	0.8993
Router semantic similarity Spearman	0.9370

MI300X serving profile

Metric	Result
vLLM / Sentence Transformers minimum cosine	0.999920
Batch-64 throughput	423.11 sequences/s
Batch-64 p50 latency	151.65 ms
32K document chunk throughput	193.44 chunks/s
Peak HBM under the frozen profile	146.32 GiB

The profile uses BF16 vLLM pooling on one AMD Instinct MI300X. Throughput is a capacity reference, not a guarantee for every server configuration.

Standard embeddings

python

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Xunzhuo/vsr-embeddings-v1-text-large", trust_remote_code=True)
vectors = model.encode(
    ["Instruct: Retrieve semantically similar text\nQuery: route this request"],
    normalize_embeddings=True,
)

vSR runtime

python

import sys
from huggingface_hub import snapshot_download

repo = snapshot_download("Xunzhuo/vsr-embeddings-v1-text-large")
sys.path.insert(0, repo)
from vsr_runtime import VSRTextEmbeddingModel

model = VSRTextEmbeddingModel(repo)

query = model.encode(
    ["Find a model that understands this request"],
    task="retrieval",
    role="query",
    dimension=1024,
)

scores = model.score_long_documents(
    "multilingual routing evidence",
    ["A long candidate document ..."],
)

Deployment guidance

Use text-small for always-on routing, semantic cache lookup, and high-QPS retrieval.
Use text-large when multilingual accuracy and difficult semantic matching dominate latency.
Direct encoding defaults to 4096 tokens. Use the included max-sim runtime for long documents instead of treating 32K text as one vector.
This is a bi-encoder embedding model, not a cross-encoder reranker and not a sequence-level safety classifier.

Foundation and license

v1 preserves the foundation geometry of Qwen/Qwen3-Embedding-4B at the revision recorded in vsr_family_manifest.json. vSR adds the router-native task, representation, and serving contract described above. See THIRD_PARTY_NOTICES.md for concise attribution. License: apache-2.0.