MB Model Series
The MB series is MUXBITE's family of generative language models — each built to be safer, more capable, and more accessible than the last.
First generation · 2025–2026
Our foundation model. A small, efficient transformer that proves the architecture works — and begins to answer the question of what MUXBITE AI feels like.
Text generation
Coherent sentence and paragraph generation in English and Hindi. Suitable for short-form writing tasks.
Basic conversation
Turn-based dialogue with short context retention. Handles simple Q&A and assistant interactions.
Translation (EN ↔ HI)
English-Hindi and Hindi-English translation at a basic level, trained on bilingual corpora.
Reasoning
Multi-step reasoning is limited at this scale. MB2 is designed to address this.
Vision
Not available in MB1. Multimodal capabilities are planned for MB3.
Tool use
Function calling and tool use are not available in MB1.
| Benchmark | Category | MB1 score | Progress |
|---|---|---|---|
| HellaSwag | Commonsense | Pending | |
| MMLU | Knowledge | Pending | |
| HumanEval | Code | Pending | |
| Indic NLU | Hindi NLU | Pending |
Benchmarks will be published when MB1 training and evaluation complete.
Second generation · Planned 2026
A more capable model with a focus on multilingual understanding, longer context, and stronger reasoning — built to serve the underserved languages of the world.
Multi-step reasoning
Chain-of-thought reasoning for math, logic, and analysis tasks across 20+ languages.
Multilingual NLU
Native understanding across Hindi, Arabic, Swahili, Bengali, Tamil, and more underserved languages.
Long-form writing
Essays, reports, stories, and documentation with a 64K token context window.
Code generation
Python, JavaScript, and other common languages. Sufficient for real-world coding assistance.
Summarisation
Long document summarisation, cross-lingual summarisation, and meeting notes.
Vision
Not available in MB2. Image understanding is planned for MB3.
| Benchmark | Category | Target | Projection |
|---|---|---|---|
| MMLU | Knowledge | ~65% | |
| HellaSwag | Commonsense | ~80% | |
| HumanEval | Code | ~40% | |
| Indic NLU | Hindi/Indic NLU | ~72% | |
| FLORES-200 | Translation | ~38 BLEU |
Targets based on scaling laws and similar models at this parameter range. Actuals will be published post-training.
Third generation · Roadmap 2027+
Our frontier model. Multimodal, multilingual, and built to compete at the highest level — text, voice, and vision for the world's next four billion users.
Advanced reasoning
Expert-level multi-step reasoning, mathematics, and scientific problem solving.
Vision understanding
Image analysis, chart reading, document OCR, and visual Q&A across languages.
Voice (speech)
Spoken language understanding and generation in 100+ languages and dialects.
Tool use & agents
Function calling, browser use, and multi-step agentic task execution.
Expert-level code
Full software engineering — architecture, debugging, code review, and security analysis.
100+ languages
Native-quality text, translation, and understanding across the full breadth of global languages.
| Benchmark | Category | Target | Projection |
|---|---|---|---|
| MMLU | Knowledge | ~85% | |
| HumanEval | Code | ~80% | |
| MATH | Mathematics | ~70% | |
| MMMU | Multimodal | ~75% | |
| FLORES-200 | Translation | ~50 BLEU |
Long-range projections based on current scaling trends. Targets may be revised as the field evolves.
| Feature | MB1 | MB2 | MB3 |
|---|---|---|---|
| Parameters | 10–50M | 200M–1B | 7B–70B+ |
| Context window | 8K | 64K | 200K |
| Languages | 2 | 20+ | 100+ |
| Text generation | ✓ | ✓ | ✓ |
| Reasoning | — | ✓ | ✓ |
| Code generation | — | ✓ | ✓ |
| Vision | — | — | ✓ |
| Voice | — | — | ✓ |
| Tool use | — | soon | ✓ |
| API access | soon | 2026 | 2027+ |
| Open weights | ✓ Planned | ✓ Planned | TBD |
| Edge deployment | ✓ | Limited | — |