Models Research About Stay updated

First generation · 2025–2026

MB1

Our foundation model. A small, efficient transformer that proves the architecture works — and begins to answer the question of what MUXBITE AI feels like.

Status In development
Parameters 10M – 50M
Context window 8K tokens
Languages English · Hindi
Modality Text
Training compute ~10²² FLOPs
Architecture Transformer (GPT-style)

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.

Context window 8K tokens
Est. throughput ~800 tok/s
Memory footprint ~200 MB
Latency (TTFT) ~50 ms
Research prototypeValidate training pipelines and architecture decisions at low cost.
Edge deploymentRun on low-resource devices — phones, embedded systems, offline apps.
Bilingual assistantSimple English-Hindi Q&A for consumer applications.
Text classificationSentiment analysis, intent detection, topic labelling.
Long-form writingNot well-suited — limited context window and scale.
Complex reasoningNot recommended — MB2 or MB3 are better suited.

Second generation · Planned 2026

MB2

A more capable model with a focus on multilingual understanding, longer context, and stronger reasoning — built to serve the underserved languages of the world.

Status Planned · 2026
Parameters 200M – 1B
Context window 64K tokens
Languages 20+ languages
Modality Text
Training compute ~10²³ FLOPs
Architecture Transformer + SFT + DPO

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.

BenchmarkCategoryTargetProjection
MMLUKnowledge~65%
65%
HellaSwagCommonsense~80%
80%
HumanEvalCode~40%
40%
Indic NLUHindi/Indic NLU~72%
72%
FLORES-200Translation~38 BLEU
est.

Targets based on scaling laws and similar models at this parameter range. Actuals will be published post-training.

Context window64K tokens
Est. throughput~200 tok/s
Memory footprint~4 GB
Latency (TTFT)~200 ms
Multilingual assistantCustomer support, Q&A, and conversation in 20+ world languages.
Document analysisSummarise and extract from long PDFs, contracts, and research papers.
Code assistantAutocomplete, debugging, and explanation for developers worldwide.
EducationPersonalised tutoring and explanations in the student's native language.
Content localisationTranslate and adapt content across global markets natively.
Image tasksNot available — use MB3 for vision capabilities.

Third generation · Roadmap 2027+

MB3

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.

StatusRoadmap 2027+
Parameters7B – 70B+
Context window200K tokens
Languages100+ languages
ModalityText · Vision · Voice
Training compute~10²⁵ FLOPs
ArchitectureMultimodal transformer

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.

BenchmarkCategoryTargetProjection
MMLUKnowledge~85%
85%
HumanEvalCode~80%
80%
MATHMathematics~70%
70%
MMMUMultimodal~75%
75%
FLORES-200Translation~50 BLEU
est.

Long-range projections based on current scaling trends. Targets may be revised as the field evolves.

Context window200K tokens
Est. throughput~80 tok/s
Memory footprint~40 GB
Latency (TTFT)~500 ms
AI assistantGeneral-purpose assistant on par with the best models in the world.
Software engineeringEnd-to-end coding, architecture review, and automated testing.
Medical & legalDomain-specific reasoning in healthcare and legal contexts globally.
Creative & mediaWriting, image understanding, and multimedia content creation.
Enterprise agentsAutonomous multi-step agents for complex business workflows.
Global educationPersonalised learning in any language, with voice and vision support.

Compare models

Feature MB1 MB2 MB3
Parameters10–50M200M–1B7B–70B+
Context window8K64K200K
Languages220+100+
Text generation
Reasoning
Code generation
Vision
Voice
Tool usesoon
API accesssoon20262027+
Open weights✓ Planned✓ PlannedTBD
Edge deploymentLimited