MenteE AI · Blog

Blog

Product updates, research notes and building in public — from the MenteE AI team. Read the technical report and model card for mentee-embed-v1.

FeaturedAugust 29, 2026 · 6 min read

Introducing MenteE SWE: An Autonomous Coding Agent for Your Terminal

MenteE SWE is a model-agnostic autonomous software-engineering agent. It investigates, edits, verifies with your tests, and reports with evidence. Install via npm, bring your own API key.

ProductMenteE SWEDeveloper Tools
Read article →

Latest articles

15 articles
GuideSep 1, 2026 · 6 min read

Migrating from mentee-embed-v3 to v4: A Practical Guide

v4 is a drop-in replacement for v3 — same dimension, same API. But re-embedding your corpus, re-tuning thresholds, and a few gotchas are worth planning for.

Guidementee-embedMigration
ResearchSep 3, 2026 · 7 min read

Hard Negative Mining: The Highest-Leverage Trick in Multilingual Embedding Training

Random in-batch negatives get you 80% of the way. The last 20% — dialects, code-switching, near-duplicate retrieval — comes from mining negatives that are almost right.

ResearchTrainingContrastive Learning
ResearchAug 28, 2026 · 8 min read

mentee-embed-v4: 146% Custom Bench Improvement, 18K sents/sec, mMARCO Arabic

MenteE AI publishes mentee-embed-v4 — 2.6M triplets, 3-round distillation, mMARCO Arabic data. Custom bench MRR@10 0.252 (+146% vs v3), MIRACL AR 0.874, 18,115 sents/sec on RTX 5090.

Researchmentee-embedAI
GuideAug 27, 2026 · 5 min read

Using MenteE SWE in CI/CD: Headless Mode for Automated Code Fixes

MenteE SWE supports headless mode for CI/CD pipelines. Run tasks without the terminal UI, get structured output, and integrate autonomous code fixes into your workflow.

GuideMenteE SWEDevOps
ResearchAug 26, 2026 · 7 min read

Arabic Embedding Benchmark 2026: How mentee-embed-v4 Approaches Pretrained Performance

mentee-embed-v4 scores 0.874 MRR@10 on MIRACL Arabic — close to mpnet-base (0.898) at 6.8× fewer parameters. A deep dive into Arabic retrieval from scratch.

ResearchArabic NLPEmbeddings
ResearchAug 25, 2026 · 6 min read

Why We Train From Scratch Instead of Fine-Tuning: The MenteE Embed Philosophy

Most embedding models fine-tune a pretrained backbone. MenteE Embed trains from random initialization. Here is why we chose the harder path and what it gives us.

ResearchAIPhilosophy
ResearchAug 26, 2026 · 9 min read

Introducing mentee-embed-v3: How Far Can Multilingual Embeddings Be Trained from Scratch?

MenteE AI publishes mentee-embed-v3 — a 41M trilingual embedding model for Arabic, English and Urdu trained from random init with 2.1M triplets, MS-MARCO retrieval data, hard negatives and batch size 512. Protocol A avg MRR@10 0.655. Published at doi:10.5281/zenodo.22117673.

Researchmentee-embedAI
ResearchAug 27, 2026 · 7 min read

Batch Size Beats Model Size: The Lesson from Training mentee-embed-v2 and v3

MenteE AI scaled up to 125M parameters for v2 and performance dropped. Returning to 41M with batch 512 produced a 3× Protocol C improvement. Here is exactly why batch size dominates over parameter count in from-scratch contrastive embedding training.

ResearchAIEmbeddings
GuideAug 28, 2026 · 8 min read

mentee-embed-v3 vs MiniLM vs mpnet: Arabic, English and Urdu Benchmark 2026

Side-by-side benchmark: mentee-embed-v3 (41M, random init) vs all-MiniLM-L6-v2, paraphrase-MiniLM-L12-v2 and mpnet-base-v2 across in-batch retrieval, Wikipedia corpus and MS-MARCO. The only sub-50M model that works across all three languages.

ResearchGuideEmbeddings
ResearchAug 24, 2026 · 7 min read

Introducing mentee-embed-v1: Training Competitive Multilingual Embeddings from Scratch for Arabic, English and Urdu

MenteE AI presents mentee-embed-v1 — a 41M-parameter trilingual embedding model for Arabic, English and Urdu trained entirely from scratch with relational knowledge distillation. Published at doi:10.5281/zenodo.22087139. Open weights, open code, honest benchmarks.

Researchmentee-embedAI
VisionAug 20, 2026 · 6 min read

Why Urdu and Arabic Matter: Building AI for 300M+ Speakers the Industry Ignores

Urdu has 230M speakers yet almost no competitive open embedding model. Arabic is thin in the small-model tier. Why MenteE AI at menteeai.org, founded by Syed Syab Ahmad (syab.tech), treats Urdu and Arabic as first-class languages.

Urdu NLPArabic NLPMenteE AI
BusinessAug 18, 2026 · 5 min read

MenteE AI Is a Product Company, Not an Agency: We Build AI Platforms That Ship

MenteE AI (menteeai.org) builds and deploys AI-powered platforms — we do not take briefs. How MenteE operates as a product lab, with open research like mentee-embed-v1 to prove it.

BusinessMenteE AIProducts
ResearchAug 16, 2026 · 8 min read

How Relational Knowledge Distillation Makes a 41M Model Beat 100M+ Giants

Contrastive training from scratch collapses. Relational distillation from multilingual-e5-base fixes it: a 41M mentee-embed student learns 960 dense numbers per batch to reach 0.585 MRR@10 vs 0.449 for all-MiniLM-L6-v2.

ResearchAIEmbeddings
CultureAug 14, 2026 · 5 min read

Building MenteE AI in 2026: From Idea to Product Lab at menteeai.org

How MenteE AI at menteeai.org became a product lab in 2026 — building mentee-embed-v1, shipping AI platforms, and publishing open research at Zenodo.

CultureMenteE AIBuilding in Public
GuideAug 12, 2026 · 9 min read

Choosing an Embedding Model in 2026: mentee-embed vs MiniLM vs E5 vs BGE Compared

A 2026 guide to embedding models for Arabic, English and Urdu: compare mentee-embed-v1 (41M) vs all-MiniLM-L6-v2, paraphrase-MiniLM-L12-v2, mpnet-base-v2, E5 and BGE — benchmarks, size and when to use each.

GuideEmbeddingsSEO