@AI_AlphaEdge

French AI lab. We design Efficient Language Models (ELM): the right model for each task, at the right cost, on your infrastructure.

 Paris, France
alphaedge-ai.com
Joined April 2026
  • Tweets20
  • Following1
  • Followers11
8Media
Hi X 👋 We're AlphaEdge, a French AI deeptech based in Angers. We design our own models from the ground up: compact, specialized, and built to run where your data lives (sovereign cloud, your own servers, or on-device). A quick thread on who we are and what's coming 🧵
2
3
3
152
Mercredi 28/10, à 11h, nous décortiquerons en direct un dossier KYB complexe avec Evarist. Un groupe fictif à trois pays : holding luxembourgeoise, filiale française, actionnaire suisse… et un bénéficiaire effectif qui n’apparaît dans aucun registre 👇
1
1
1
15
Nos intervenants : 👉 Théo Hubert, CTO d’AlphaEdge : AlphaIDP lit, classe et extrait chaque pièce du dossier. 👉 Virgile Augendre, CEO d’Evarist : Soma relie les pièces pour trouver qui contrôle vraiment la société. Inscription 👉 eu1.hubs.ly/H0z5r_l0
13
👀 Three OCR APIs, one complex table, focus on section IV: • DeepSeek: 35+s, only partly reproduced & filled. • Mistral: 3.98s, structure not fully preserved. • AlphaDigitMax v2: 5.9s, table reproduced & filled + reliability score. More tests soon. 🤗
2
2
17
Comment l’IA peut-elle sécuriser l’analyse de gouvernance et réduire le Time-to-Revenue ? Le 28/10 à 11h, avec Evarist, nous répondons en direct sur un cas KYB réaliste : 45 min de démo et de questions. Pour les équipes conformité, KYB/KYC et risques, en banque et assurance. 🧵
2
2
2
25
Au programme : 📄 Extraire automatiquement la donnée des documents de gouvernance avec AlphaIDP. 🕸 Cartographier les liens capitalistiques en langage naturel avec Soma, le GraphRAG d’Evarist. ⚖️ Distinguer détention et contrôle effectif, avec une analyse traçable.
1
20
👀 Un bonus réservé aux participants présents en direct. Inscription gratuite 👉 eu1.hubs.ly/H0z5r_l0
4
SOTA sparse retrieval, now in French. We’re releasing Sparseup-fr-v0 on Hugging Face, our second open-source release this year.
1
2
39
more replies
On a cleaned NanoBEIR-fr (13 datasets): 56.37 average, the best of the open-source French sparse embedding models we evaluated. Best on 8 of 13 datasets, +7.9 points over BM25.
1
1
12
It’s a V0: our Lead Data Scientist @BdsLoick replicated @Linkup_platform’s recipe, built on @LightOnIO's work, to check it transfers to French. It does. A new version is already in the works. Model, benchmark and blog post: huggingface.co/alphaedge-ai/…

alphaedge-ai/sparseup-fr-v0 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
1
24
Yesterday was about the top score. Today, the bigger picture: all three AlphaDigit v2 models land in the top 5 of olmOCR-Bench. Even our lightest, AlphaDigitFast v2, passes 83.0% of tests, ahead of 5 of the 7 other systems evaluated. Full thread ↓
AlphaEdge@AI_AlphaEdgeOct 8
Our OCR model AlphaDigitMax v2 scores 89.4% on olmOCR-Bench (@allen_ai): the best overall score among the 10 systems we evaluated, ahead of Chandra (86.9%) and Mistral OCR (83.8%). All three AlphaDigit v2 models rank in the top 5 🧵
33
Our OCR model AlphaDigitMax v2 scores 89.4% on olmOCR-Bench (@allen_ai): the best overall score among the 10 systems we evaluated, ahead of Chandra (86.9%) and Mistral OCR (83.8%). All three AlphaDigit v2 models rank in the top 5 🧵
1
1
1
1
122
more replies
Full report (protocol, per-category results, list of corrections): bit.ly/4ennNG3 Models on Hugging Face: huggingface.co/alphaedge-ai

AlphaDigitMaxv2-En.pdf

drive.google.com
16
Overall scores on 1,402 PDF pages and ~8,400 pass/fail tests. The biggest gap is on the hardest category for every system, old & degraded scans: 60.6% vs 51.2% for the next best (+9.4 pts). Tables: 92.9% (+3.0 pts).
1
1
44
Method, same rules for everyone: one pass per page, same judge, same normalization, local models on a single RTX PRO 6000. We audited the test set and fixed impossible tests (expected values only). Several are tests our own models still fail.
24
We showed it on speech earlier this year: AlphaAudio v0 ranked #1 on CommonVoice (French), ahead of Voxtral Transcribe v2, while running 5x faster than Whisper Large v3. Tomorrow, it's documents' turn. ♟️
1
1
3
23
📄 Tomorrow: we publish the evaluation of our OCR models, AlphaDigit v2, on a public benchmark: protocol, results by category, and our limits. Want to see it first? Our Discord gets it before anyone else 👇 discord.gg/zHW6aWFve

Join the AlphaEdge Community Discord Server!

Check out the AlphaEdge Community community on Discord - hang out with 23 other members and enjoy free voice and text chat.

discord.com
1
3
19
Load more