AsteroRank-Lepton

AsteroRank-Lepton

Scores how relevant each document is to a search query, with optional context, to reorder search results. A 4B open-weights reranker for science and general search, built on Qwen3.5.

Made by Asteromorph on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: AsteroRank-Lepton is a reranker from Asteromorph, built on Qwen/Qwen3.5-4B. You give it a query, optional context and a document, and it returns a relevance score between 0 and 1, where higher means a closer match. It is meant to reorder search results for science topics and general search.

Who it suits: Developers of search, retrieval and research-agent systems who want to reorder candidate documents by relevance before showing them or passing them to a model.

Industry | Who uses it | How they use it | Example
Software | Search engineers | Rerank candidate documents after a first retrieval step | Reordering ten hits for a query
Research | Scientists and lab teams | Find the paper passages that match a technical question | Matching a chemistry query to abstracts
AI agents | Agent builders | Let a search agent pass along context that sharpens ranking | A research agent that supplies a short summary
Education | Teaching teams | Surface the best study material for a topic | Sorting readings for a course question

If you only need first-pass keyword search, or need commercial use, you can skip it, since the license is non-commercial.

Where to use it:

• SGLang, vLLM and Sentence Transformers, which the card says work from day one
• Windows, macOS or Linux with Python, for downloading the files
• Search and retrieval pipelines that call a rerank endpoint

Use case | Who | What they get back
Reranking results | Search engineers | Documents ordered from most to least relevant
Context-aware ranking | Agent builders | Scores that use an extra context passage
Query-only ranking | Developers | Scores based on the query alone when context is left out

How to set it up:

• Python, to run the Hugging Face download tool
• A GPU able to serve a 4.8B model, where the card's own test used one NVIDIA B200
• SGLang, vLLM or Sentence Transformers to serve the model

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Start a server with SGLang, vLLM or Sentence Transformers.
4. Send the query, optional context and candidate documents to the rerank endpoint.

pip install -U huggingface_hub
hf download Asteromorph/AsteroRank-Lepton

How to use it: Start the model with SGLang using the card's tested command, then send rerank requests to it. The card keeps Qwen3.5's standard configuration, so SGLang needs no code changes.

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True sglang serve \
  --model-path Asteromorph/AsteroRank-Lepton \
  --port 30000 \
  --attention-backend trtllm_mha \
  --mem-fraction-static 0.6 \
  --max-running-requests 64 \
  --max-mamba-cache-size 1280 \
  --cuda-graph-max-bs-decode 64 \
  --chunked-prefill-size 65536 \
  --max-prefill-tokens 65536 \
  --disable-prefill-cuda-graph

Tool | What it does
query | The search text you want documents ranked against
instruct | Optional context that helps the ranking
documents | The candidate texts to score
top_n | Limits how many ranked documents come back

Example: You send a chemistry query about para-selective borylation, a short context passage and three candidate documents. The server returns scores of 0.6792 for the second document, 0.2814 for the first and 0.0474 for the third, so the closest match comes first.

For newcomers and experts: Start with one query and three short documents, and check that the one you expect scores highest. Leave out the context field at first to see the query-only behavior.

The card says the model keeps Qwen3.5's standard configuration, and vLLM serves it as a one-label classifier through its pooling runner. Giving the model a context passage is called context reranking in the card. The family's blog post also explains how to prompt a search agent so it writes useful context.

Send query, context and documents
Model scores each document
Scores fall between 0 and 1
Documents come back ranked

Good to know:

• License: CC BY-NC 4.0, which is non-commercial
• Cost to run: none named, though the card's test used a high-end GPU
• Data it reaches: downloads files from Hugging Face, then runs where you serve it
• Not for: commercial use under the stated license
• Not for: mixing rerank requests with chat or generate calls on the same SGLang server, since the card says that crashes the scheduler
• Limits: keep the chunked prefill size below 174,764, which the card gives as the safe bound

On Hugging Face, 13 November 2026

• Task: text ranking
• Library: transformers
• Parameters: 4.8B
• Weights: about 9.7 GB in BF16; about 2.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 58
• Likes: 7
• License: cc-by-nc-4.0
• Languages: en
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 2
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, reranker, sglang, sentence-transformers, text-ranking

Links

• Model: https://huggingface.co/Asteromorph/AsteroRank-Lepton
• Files: https://huggingface.co/Asteromorph/AsteroRank-Lepton/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Asteromorph/AsteroRank-Lepton

Then follow the usage in its model card. Nothing in this kit runs by itself.