Explore our foundational research papers, model architectures, technical benchmarks, and latest company releases from MBiG.INc.
We introduce DThink-1, a reasoning model engineered via large-scale self-supervised pre-training and outcome-supervised reward modeling. DThink-1 achieves competitive breakthroughs across MATH-500, Codeforces, and complex logic benchmarks.
Authors: MBiG AI Research Team (MBiG.INc)
The official release of MBiG AI 3.1 Flash features enhanced token generation throughput, sub-second latency for Indian multilingual prompts, and unified multimodal image and document understanding.
A technical deep dive into MBiG's custom Indic subword tokenizer, designed to eliminate token fragmentation in Hindi, Hinglish, Marathi, and Tamil texts while significantly accelerating generation speeds.
Authors: MBiG NLP Lab
Comprehensive benchmark evaluations comparing MBiG AI engines against industry baselines. Highlights include a 91.4% pass@1 on GSM8k with DThink enabled.
How MBiG AI combines dynamic search query generation, Multi-Source Fact Checking, and real-time neural reranking to deliver verifiable citations in chat responses.
Authors: MBiG Systems Engineering
Announcing public availability of MBiG REST APIs and SDKs with OpenAI-compatible endpoints, structured JSON outputs, and enterprise SLA support.
Get early announcements on model architecture disclosures, benchmark evaluations, and AI releases straight to your inbox.