README.md

---
license: other
license_name: proprietary
license_link: https://voxell.ai/forge
base_model: Qwen/Qwen3-Embedding-8B
pipeline_tag: feature-extraction
library_name: sentence-transformers
language:
- en
tags:
- sentence-transformers
- feature-extraction
- sentence-similarity
- mteb
- qwen3
- qwen
- text-embedding
arxiv: 2502.13595
---

# Ingot-8B-R3

**An 8B English text-embedding model with per-task instruction routing and per-cluster
specialists, served as an API.** Built by [Jonathan Corners](https://voxell.ai) at
**[Voxell](https://voxell.ai)**.

- 🏠 **Home:** https://voxell.ai
- 📝 **Engineering write-up — *Ingot Poured*:** https://voxell.ai/engineering/ingot_poured/
- 🧪 **Technical essays:** https://sentimark.ai
- 🔥 **Use it — 10M tokens free, then pay-as-you-go / select a plan:** https://voxell.ai/forge
- 🎮 **Playground:** https://playground.voxell.ai/
- 💼 **LinkedIn:** [Jonathan Corners](https://www.linkedin.com/in/jonathancorners/) · [Voxell](https://www.linkedin.com/company/voxell-ai/)
- ✉️ **Contact:** corp@voxell.ai

> **Access model.** This is a **proprietary, API-only** model — no weights are distributed here.
> **To use it, sign up at [voxell.ai/forge](https://voxell.ai/forge) — 10M tokens free, then pay-as-you-go / select a plan.**
> MTEB reviewers can get a free per-reviewer key to reproduce the benchmark — see
> [Access & usage](#access--usage) below.

---

## Overview

| | |
|---|---|
| **Base model** | [`Qwen/Qwen3-Embedding-8B`](https://huggingface.co/Qwen/Qwen3-Embedding-8B) (rev `1d8ad4ca…`) |
| **Embedding dimension** | 4096 (float32) |
| **Max sequence length** | 32,768 tokens |
| **Similarity** | Cosine |
| **Language** | English (`eng-Latn`) |
| **Parameters** | 7.57B total · 621M embedding |
| **Serving** | OpenAI-compatible `/v1/embeddings` |

Ingot-8B-R3 wraps the Qwen3-Embedding-8B backbone with a routing framework that selects
per-task instructions and activates per-cluster specialist experts behind a single
embeddings endpoint. The routing logic is proprietary (provisional patent filed 2026)
and is not included here; the model is consumed through the API. Full method notes are in
the engineering write-up: **[*Ingot Poured*](https://voxell.ai/engineering/ingot_poured/)**.

## Results — MTEB(eng, v2)

| Metric | Score |
|---|---|
| **Mean (Task)** | **75.9795** |
| Mean (Category) | 69.9958 |
| Borda points | 5567 |

| Category | Score |
|---|---|
| Classification | 90.41 |
| STS | 89.32 |
| PairClassification | 87.66 |
| Retrieval | 70.01 |
| Clustering | 58.47 |
| Summarization | 36.96 |
| Reranking | 32.84 |

At evaluation time this was, to our knowledge, the highest **Mean (Task)** on MTEB(eng, v2)
among English embedding models developed in the United States. Built by a single engineer on
consumer GPUs — the emphasis is data engineering over compute scale (see the write-up).

Results are submitted to [`embeddings-benchmark/results`](https://github.com/embeddings-benchmark/results)
and the registry PR to [`embeddings-benchmark/mteb`](https://github.com/embeddings-benchmark/mteb)
is open for review.

## Access & usage

### Production use → [voxell.ai/forge](https://voxell.ai/forge)

For any real / production use, get an API key at **[voxell.ai/forge](https://voxell.ai/forge)**:
**10M tokens free**, then pay-as-you-go or select a plan. OpenAI-compatible endpoint, no weights to host.

### MTEB benchmark reproduction

For reviewers reproducing MTEB(eng, v2), a separate audited eval endpoint is served at
**`https://api-mteb.voxell.ai`** with an OpenAI-compatible body shape.

**1. Get a key** (~30 s, self-serve, no email round-trip):
Visit **https://api-mteb.voxell.ai/request-key**, enter your email, complete the
Cloudflare challenge, and your key appears on screen with a prefilled `curl` example.
200 requests/minute per reviewer.

**2. Check status without a key:**

```bash
curl -sS https://api-mteb.voxell.ai/v1/models
# → {"object":"list","data":[{"id":"jcorners/ingot-8b-r3","object":"model","owned_by":"voxell"}]}
```

**3. Embed:**

```python
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MTEB_API_KEY"],
    base_url="https://api-mteb.voxell.ai/v1",
)

resp = client.embeddings.create(
    model="jcorners/ingot-8b-r3",
    input=["Example sentence"],
)
print(len(resp.data[0].embedding))  # 4096
```

Or with `curl` (note `task_name` for instruction routing):

```bash
curl -sS https://api-mteb.voxell.ai/v1/embeddings \
  -H "Authorization: Bearer $MTEB_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"jcorners/ingot-8b-r3","input":["hello world"],"task_name":"STS12"}'
```

### Use from MTEB

The model registers as `jcorners/ingot-8b-r3` via an API loader
(`mteb/models/model_implementations/ingot_models.py`). Set `MTEB_API_KEY` and run MTEB normally.

## Training data disclosure

`training_datasets` is declared in full on the ModelMeta — two groups: datasets inherited
from the Qwen3-Embedding-8B base, and the MTEB train splits used by Ingot R3 specialists
(SprintDuplicateQuestions, TwitterSemEval2015, TwitterURLCorpus, MedrxivClusteringP2P.v2,
MedrxivClusteringS2S.v2). Ingot's own synthetic training data is generated through a
source-document-level leakage gate, so MTEB eval-set documents are quarantined by construction.

## Citation

If you use this model, please reference the engineering write-up
[*Ingot Poured*](https://voxell.ai/engineering/ingot_poured/) and Voxell (https://voxell.ai).

The MTEB benchmark: [MMTEB: Massive Multilingual Text Embedding Benchmark](https://arxiv.org/abs/2502.13595).