Anshulgada/tabman-embeddinggemma-2-440m

Tabman EmbeddingGemma 2 (440M)

Tabman EmbeddingGemma 2 (440m) is a state-of-the-art sentence and document embedding model fine-tuned specifically for browser tab organization, semantic search, hierarchical URL clustering, and web history discovery in Tabman (Brave Tab Manager).

Trained on 18,000 stratified contrastive triplets across 18 distinct web verticals using Matryoshka Representation Learning (MRL) with loss penalties computed simultaneously at 768d, 512d, 256d, and 128d.

---

Model Specifications

Parameter	Specification
Model ID	Anshulgada/tabman-embeddinggemma-2-440m
Base Architecture	google/embeddinggemma-2 (~440M parameters)
Modality Scope	Multimodal Text + Vision transformer backbone (~440M parameters)
Native Dimension	768 dimensions
Matryoshka Dimensions	768d, 512d, 256d, 128d (Recommended)
Max Context Window	8,192 tokens (512 tokens tested for tab titles & URLs)
Training Precision	Pure bfloat16 (bf16) on NVIDIA RTX A6000 (48GB VRAM)
Loss Function	MultipleNegativesRankingLoss wrapped in MatryoshkaLoss
Base Checkpoint	Anshulgada/embeddinggemma-2-440m
License	Apache 2.0

---

The Matryoshka Miracle: 6x Smaller with Zero Quality Loss

Traditional embedding models require storing 768 or 1024 floating-point values per tab or document. With Matryoshka Representation Learning (MRL), front-truncated slices of the embedding vector preserve near-identical semantic clustering geometry:

Output Dimension	Accuracy (%)	Mean Cosine Margin	Vector Size Ratio	Vector RAM / 10,000 Tabs
768d (Native)	98.50%	+0.6074	1.00x (Baseline)	~30.7 MB
512d	98.50%	+0.6077	0.67x	~20.5 MB
256d	98.44%	+0.6051	0.33x	~10.2 MB
128d (Target Runtime)	98.67%	+0.6427	0.17x (6x smaller!)	~5.1 MB

Truncating embeddings from 768d to 128d (embedding[:128]) results in 0.00% accuracy loss across the entire 1,800 test triplet suite while cutting browser memory, index storage, and vector distance computation time by 83.3%.

---

Empirical Benchmark Findings across 18 Domains

Evaluated across 1,800 held-out test triplets stratified across 18 real-world web domains:

• 100.0% Accuracy on 17 out of 18 Domains:
  • API Documentation (MDN, Stripe, PyTorch)
  • Cloud & DevOps (AWS, GCP, Kubernetes, Terraform)
  • Frontend Development (React, Vue, Tailwind, Next.js)
  • Backend & Databases (PostgreSQL, Redis, Kafka, FastAPI)
  • Machine Learning & AI (Hugging Face, PyTorch, LangChain, CUDA)
  • Productivity & Workspace (Notion, Google Docs, Slack, Jira)
  • Localhost & Internal Dev (localhost:3000, Swagger, Grafana)
  • Community & Forums (StackOverflow, Reddit, GitHub Discussions)
  • Career & Job Search (LinkedIn, Greenhouse, Indeed)
  • Design & UI/UX (Figma, Dribbble, Unsplash)
  • Finance & Banking (Stripe, Robinhood, Bloomberg)
  • Video Streaming (YouTube, Twitch, Netflix)
  • Audio & Podcasts (Spotify, SoundCloud, Apple Podcasts)
  • E-Commerce & Shopping (Amazon, eBay, Shopify)
  • News & Journalism (NYTimes, BBC, TechCrunch)
  • Travel & Booking (Airbnb, Google Flights, Booking.com)
  • Gaming & Esports (Steam, Discord, IGN)
• 72.7% - 75.8% Accuracy on Dense Academic Research:
  • ArXiv CS/Math dense papers with overlapping terminology.

---

Sizing & Deployment Matrix

Runtime Environment	Format / Engine	Memory / VRAM	Latency / Tab	Throughput	Best For
Browser Extension	WebGPU INT8 ONNX (Transformers.js)	~270 MB	8 - 15 ms	~80 tabs/sec	Real-time tab grouping in browser
Local CPU Fallback	INT8 ONNX (onnxruntime)	~280 MB	30 - 45 ms	~25 tabs/sec	Systems without WebGPU / discrete GPU
Remote / LAN Server	PyTorch CUDA BF16 (serve_embeddings.py)	~600 MB	< 2 ms	>1,500 tabs/sec	Centralized indexing across multi-browser setups

---

Quickstart & Code Examples

1. Python (sentence-transformers) with 128d Slicing

from sentence_transformers import SentenceTransformer

# Load model (automatically uses CUDA if available)
model = SentenceTransformer("Anshulgada/tabman-embeddinggemma-2-440m")

queries = [
    "GitHub PR #24502: Fix memory leak in reconciliation queue",
    "Stack Overflow: Concurrency race condition in async runtime",
    "Stripe API: Handling Webhook idempotency keys in Python",
]

# Generate normalized 128-dimensional Matryoshka embeddings
embeddings = model.encode(queries, normalize_embeddings=True)
embeddings_128d = embeddings[:, :128]

similarity = model.similarity(embeddings_128d[0], embeddings_128d[1])
print(f"Cosine Similarity (PR vs StackOverflow): {similarity.item():.4f}")

2. In-Browser WebGPU via Transformers.js (JavaScript / TypeScript)

This repository includes pre-converted, quantized INT8 ONNX weights in the onnx/ folder, ready for direct execution inside Chrome / Brave extensions:

import { pipeline } from "@huggingface/transformers";

// Initialize in-browser feature extraction with WebGPU acceleration
const extractor = await pipeline(
  "feature-extraction",
  "Anshulgada/tabman-embeddinggemma-2-440m",
  { dtype: "q8", device: "webgpu" }
);

// Extract embedding for a browser tab
const output = await extractor(
  "GitHub - Anshulgada/brave-tab-manager: AI tab discovery and manager",
  { pooling: "mean", normalize: true }
);

// Slice to 128 dimensions for ultra-fast vector clustering
const embedding128 = Array.from(output.data).slice(0, 128);
console.log("128d WebGPU Vector:", embedding128);

3. OpenAI-Compatible Remote REST API

Serve this model with the included zero-dependency server:

python scripts/ai/serve_embeddings.py --port 11434 --variant 440m

Consume using standard OpenAI SDK:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:11434/v1", api_key="not-needed")
response = client.embeddings.create(
    input=["GitHub PR #42: Brave Tab Manager", "Kubernetes cluster configuration"],
    model="Anshulgada/tabman-embeddinggemma-2-440m",
)
print("Embedding dimension:", len(response.data[0].embedding))  # 128

4. Ollama-Compatible REST API

curl http://localhost:11434/api/embeddings \
  -H "Content-Type: application/json" \
  -d '{"model": "Anshulgada/tabman-embeddinggemma-2-440m", "prompt": "MDN Web Docs: WebGPU API fundamentals"}'

5. Direct ONNX Runtime (Python / C++ / Rust)

import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

session = ort.InferenceSession("onnx/model_quantized.onnx")
tokenizer = AutoTokenizer.from_pretrained("Anshulgada/tabman-embeddinggemma-2-440m")

inputs = tokenizer(["Brave Browser: Tab Groups API"], return_tensors="np")
ort_inputs = {
    "input_ids": inputs["input_ids"],
    "attention_mask": inputs["attention_mask"],
}
outputs = session.run(None, ort_inputs)
raw_vec = outputs[0][0]
vec_128 = raw_vec[:128] / np.linalg.norm(raw_vec[:128])

---

Training Hyperparameters

• Base Architecture: google/embeddinggemma-2
• Training Epochs: 3
• Batch Size: 64 (with hard in-batch negatives)
• Loss Formulation: MultipleNegativesRankingLoss wrapped in MatryoshkaLoss([768, 512, 256, 128])
• Learning Rate: 2e-5 with 10% linear warmup
• Optimizer: AdamW (weight_decay=0.01)
• Hardware Acceleration: Pure bfloat16 on NVIDIA RTX A6000 48GB VRAM (Ubuntu 24.04 LTS)

---

The Complete Tabman AI Model Matrix

All 6 models (3 raw base models + 3 domain-specialized fine-tuned models) are published under Apache 2.0:

Model Variant	Base Model (Raw Foundation)	Fine-Tuned Model (Tabman + WebGPU)	Parameters	Focus Area
270M	Anshulgada/embeddinggemma-2-270m	Anshulgada/tabman-embeddinggemma-2-270m	~270M	Text & Code / In-Browser Extension
440M	Anshulgada/embeddinggemma-2-440m	Anshulgada/tabman-embeddinggemma-2-440m	~440M	Text + Vision Multimodal
740M	Anshulgada/embeddinggemma-2-740m	Anshulgada/tabman-embeddinggemma-2-740m	~740M	Full Multimodal (Text + Vision + Audio)

---

Citation & Project Reference

@software{gada2026tabman,
  author = {Anshul Gada},
  title = {Tabman: AI-Powered Tab Manager and Semantic Discovery},
  year = {2026},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/Anshulgada/brave-tab-manager}}
}

Developed with ❤️ for Brave Tab Manager (Tabman).