MathLeap-Qwen-8B by University of Washington Math AI Lab

Runs On

What it takes to serve MathLeap-Qwen-8B (7.6B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

Precision	Weights	Memory needed	Cheapest setup	Per hour	Also fits
16-bit	15.1 GB	18.2 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
8-bit	7.6 GB	9.1 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
4-bit	3.8 GB	4.5 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 7, 2026.

MathLeap-Qwen-8B on every accelerator the SAVRN Index prices, at every precision

Model Card

By University of Washington Math AI Lab, published under apache-2.0, revision ffe9c2a4a1b7.

A retrieval-tuned embedding model for mathematical text. Fine-tuned from Qwen/Qwen3-Embedding-8B on mathlib4 concepts via multi-view contrastive learning.

On MELD — a benchmark of mathematical statements paired across radically different presentations (e.g., set-theoretic vs. category-theoretic phrasings of the same theorem) — MathLeap-Qwen-8B achieves MMR 0.43, beating its base Qwen3-Embedding-8B (0.32, +0.11) and the retrieval-specialized Octen-Embedding-8B (0.42).

Model	Base	AMP MMR ↑ (specialized prompt)
Qwen3-Embedding-8B	—	0.32
Octen-Embedding-8B	Qwen3-8B	0.42
MathLeap-Qwen-8B (this)	Qwen3-8B	0.43

Usage

from huggingface_hub import snapshot_download
from sentence_transformers import SentenceTransformer

# Download model files from anonymous mirror
model_path = snapshot_download(
    repo_id="anonymous-submission/MathLeap-Qwen-8B",
    endpoint="https://anonymous-hf.up.railway.app/a/pv25ongyl2qb/ ", 
)

# Load locally
model = SentenceTransformer(model_path)

query = "For any natural number n, n + 0 = n."
docs = [
    "theorem add_zero (n : ℕ) : n + 0 = n := rfl",
    "theorem mul_zero (n : ℕ) : n * 0 = 0 := rfl",
]
q_emb = model.encode([query])
d_emb = model.encode(docs)
print(q_emb @ d_emb.T)

Model details

Read the full model card (750 words)

Configuration

Architecture: Qwen3Model
Context length (tokens): 40,960
Layers: 36
Hidden size: 4,096
Feed-forward size: 12,288
Attention heads: 32
Key/value heads: 8
Head dimension: 128
Vocabulary size: 151,665
RoPE base: 1,000,000
Model type: qwen3

Identity and Version

Repository: uw-math-ai/MathLeap-Qwen-8B
Publisher: University of Washington Math AI Lab
Task: Sentence similarity
Modality: Text
Library: sentence-transformers
Parameters: 7.6B parameters
Languages: en
Revision: ffe9c2a4a1b7777b890044e40a4e4a4d6346ae23
First published: 2026-05-25
Last updated: 2026-10-01

Files and Weights

22 files, 30.3 GB in total. The weights are 7 files totalling 30.3 GB in safetensors.

Weights7 files · 30.3 GB

Configuration8 files · 34.3 KB

Tokenizer4 files · 15.9 MB

Documentation1 file · 7.0 KB

Other1 file · 2.4 KB

Repository1 file · 1.6 KB

Every file

File	Type	Size	SHA-256
model-00001-of-00007.safetensors	Weights	5.0 GB	483abccff841
model-00002-of-00007.safetensors	Weights	4.8 GB	3db9485404f4
model-00003-of-00007.safetensors	Weights	4.8 GB	c9935c39aed4
model-00004-of-00007.safetensors	Weights	5.0 GB	4260beb650eb
model-00005-of-00007.safetensors	Weights	4.8 GB	dba20754eefb
model-00006-of-00007.safetensors	Weights	4.8 GB	b52d3cc979bd
model-00007-of-00007.safetensors	Weights	973.2 MB	6f672adf14db
1_Pooling/config.json	Configuration	313 B	—
added_tokens.json	Configuration	605 B	—
config.json	Configuration	1.5 KB	—
config_sentence_transformers.json	Configuration	375 B	—
model.safetensors.index.json	Configuration	30.5 KB	—
modules.json	Configuration	349 B	—
sentence_bert_config.json	Configuration	57 B	—
special_tokens_map.json	Configuration	613 B	—
README.md	Documentation	7.0 KB	—
chat_template.jinja	Other	2.4 KB	—
.gitattributes	Repository	1.6 KB	—
merges.txt	Tokenizer	1.7 MB	—
tokenizer.json	Tokenizer	11.4 MB	847b917cf652
tokenizer_config.json	Tokenizer	4.9 KB	—
vocab.json	Tokenizer	2.8 MB	—

License and Download

License: apache-2.0
Access: Open weights, no gate
Download size: 30.3 GB

Download from University of Washington Math AI Lab

Released by University of Washington Math AI Lab through its official repository on Hugging Face. Read the license.

Built From

Derived from Qwen/Qwen3-Embedding-8B

Memory Requirements

Precision	Weights in memory
As published	30.3 GB
16-bit	15.1 GB
8-bit	7.6 GB
4-bit	3.8 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About MathLeap-Qwen-8B

How much GPU memory does MathLeap-Qwen-8B need?

About 18.2 GB at 16-bit and 4.5 GB at 4-bit: the weights (7.6B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run MathLeap-Qwen-8B on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use MathLeap-Qwen-8B commercially?

Yes. MathLeap-Qwen-8B is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

What is MathLeap-Qwen-8B's context length?

40,960 tokens, from the maximum position embeddings in its published configuration.

Similar Models

Model · Sentence similarity

Octen-Embedding-8B

Octen-Team

Octen-Embedding-8B is a text embedding model developed by Octen for semantic search and retrieval tasks. This model is fine-tuned from Qwen/Qwen3-Embedding-8B and supports multiple languages, providing high-quality embeddings for various applications. - Octen-Embedding-8B ranks #1 on the RTEB Leaderboard with Mean (Task) score of 0.8045 - Excellent performance on both Public (0.7953) and Private (0.8157) datasets - Demonstrates true generalization capability without overfitting to public benchmarks - Supports up to 32,768 tokens context length - Suitable for processing long documents in legal, healthcare, and other domains - High-dimensional embedding space for rich semantic representation…

Open weights apache-2.0 7.6B parameters 40,960 tokens sentence-transformers

View model

Model · Sentence similarity

AX-Qwen3-Embedding-8B-CUDA-AXQ-NVFP4-W4A4

AutomatosX

AXQuant CUDA NVFP4 W4A4 mixed precision development preview. Converted from the original upstream BF16 source, without AWQ. Native E2M1 FP4 weights and inputs use per-16 E4M3FN scales and FP32 global scales. This is a retrieval embedding checkpoint, with no MTP or generative claim. All attention projections and the first/last two MLP blocks retain original BF16. Remaining MLP matrices use NVFP4 W4A4. Saved query instruction, last-token pooling, exactly one (151643), and L2 normalization. The chat (151645) is not appended. Factory conversion uses AXQuant's NumPy reference RTN encoder, independently checked against preserved BF16 source tensors. Source and pooling assets, calibration and…

Open weights apache-2.0 7.6B parameters 40,960 tokens vllm

View model

Model · Sentence similarity

AX-Nemotron-3-Embed-8B-CUDA-AXQ-NVFP4-W4A4

AutomatosX

AXQuant CUDA NVFP4 W4A4 mixed precision development preview. Converted from the original upstream BF16 source, without AWQ. Native E2M1 FP4 weights and inputs use per-16 E4M3FN scales and FP32 global scales. This is a retrieval embedding checkpoint, with no MTP or generative claim. Eligible attention and MLP matrices use NVFP4 W4A4. Embeddings and norms retain original BF16. Saved query: / passage: prefixes, bidirectional attention, mean pooling and L2 normalization. The native vLLM Ministral alias retains iscausal=false and every actual attention layer is verified as encoder-only. Factory conversion uses AXQuant's NumPy reference RTN encoder, independently checked against preserved BF16…

Open weights other 8B parameters 262,144 tokens vllm

View model

Model · Sentence similarity

Qwen3-VL-Embedding-8B

Qwen

The Qwen3-VL-Embedding and Qwen3-VL-Reranker model series are the latest additions to the Qwen family, built upon the recently open-sourced and powerful Qwen3-VL foundation model. Specifically designed for multimodal information retrieval and cross-modal understanding, this suite accepts diverse inputs including text, images, screenshots, and videos, as well as inputs containing a mixture of these modalities. While the Embedding model generates high-dimensional vectors for broad applications like retrieval and clustering, the Reranker model is engineered to refine these results, establishing a comprehensive pipeline for state-of-the-art multimodal search. Qwen3-VL-Embedding-8B has the…

Open weights apache-2.0 8.1B parameters 262,144 tokens sentence-transformers

View model

Model · Sentence similarity

AX-Qwen3-Embedding-4B-CUDA-AXQ-NVFP4-W4A4

AutomatosX

AXQuant CUDA NVFP4 W4A4 mixed precision development preview. Converted from the original upstream BF16 source, without AWQ. Native E2M1 FP4 weights and inputs use per-16 E4M3FN scales and FP32 global scales. This is a retrieval embedding checkpoint, with no MTP or generative claim. All attention projections and the first/last two MLP blocks retain original BF16. Remaining MLP matrices use NVFP4 W4A4. Saved query instruction, last-token pooling, exactly one (151643), and L2 normalization. The chat (151645) is not appended. Factory conversion uses AXQuant's NumPy reference RTN encoder, independently checked against preserved BF16 source tensors. Source and pooling assets, calibration and…

Open weights apache-2.0 4B parameters 40,960 tokens vllm

View model

Model · Sentence similarity

Qwen3-VL-Embedding-2B

Qwen

The Qwen3-VL-Embedding and Qwen3-VL-Reranker model series are the latest additions to the Qwen family, built upon the recently open-sourced and powerful Qwen3-VL foundation model. Specifically designed for multimodal information retrieval and cross-modal understanding, this suite accepts diverse inputs including text, images, screenshots, and videos, as well as inputs containing a mixture of these modalities. While the Embedding model generates high-dimensional vectors for broad applications like retrieval and clustering, the Reranker model is engineered to refine these results, establishing a comprehensive pipeline for state-of-the-art multimodal search. Qwen3-VL-Embedding-2B has the…

Open weights apache-2.0 2.1B parameters 262,144 tokens sentence-transformers

View model