AI / RESEARCHAI Model

M31Genesis

M31-Genesis-AgenticCode-v1 — native M31 causal language model for agentic coding research. Checkpoints are experimental research checkpoints.

STATUS
Experimental
STAGE
RESEARCH
VERSION
AgenticCode-v1
LICENSE
MIT
Experimental Research Notice

Experimental research checkpoints. Evaluate before production use. Stored on Hugging Face.

LIFECYCLE STATUS

Experimental

Experimental research artifact

DEVELOPMENT STAGE

RESEARCH

Active research & evaluation

CLASSIFICATION

AI Model

425M-class

AFFILIATION & LICENSE

TIV · MIT

Tonmoy Infrastructure & Vision

OVERVIEW

About M31Genesis

M31-Genesis-AgenticCode-v1 — native M31 causal language model for agentic coding research. Checkpoints are experimental research checkpoints.

THE PROBLEM

Why this system exists

Agentic coding systems often rely on massive, proprietary black-box APIs with unknown training distributions, unpredictable latencies, and high resource costs. Research into lightweight, sovereign causal architectures specifically optimized for agentic reasoning and code modification has been constrained by a lack of open, inspectable checkpoints.

OUR APPROACH

Engineering philosophy & approach

M31Genesis implements a native M31 decoder-only transformer architecture with 425M-class parameters. It balances expressiveness for code syntax and AST transformations with lightweight execution, publishing transparent step-wise checkpoint lineages on Hugging Face for reproducible evaluation.

ARCHITECTURE

System architecture

Native M31 decoder-only transformer: 425M-class parameters, 65,536 ByteLevel BPE vocabulary, 24 layers, hidden size 1024, 16 attention heads / 8 KV heads (GQA), RMSNorm normalization, Rotary Position Embeddings (RoPE), and tied embeddings.

CAPABILITIES

Core system capabilities

  • Native M31 decoder-only transformer design
  • 425M-class parameters optimized for local agent execution
  • 65,536 ByteLevel BPE vocabulary with code-token density
  • 16 query heads / 8 KV heads grouped-query attention
  • Rotary Position Embeddings (RoPE) and RMSNorm
  • Step-wise immutable checkpoint lineage
  • Open safetensors serialization and reproducible runtime
AI MODEL ARCHITECTURE & VERIFIED PARAMETERS

Technical Model Specification

Architecture Summary

Native M31 decoder-only transformer, 425M-class parameters, 65,536 ByteLevel BPE vocabulary, 24 layers, hidden size 1024, 16 attention heads / 8 KV heads, RMSNorm, RoPE, tied embeddings.

Verified Intended Use Cases
✓Software engineering
✓Coding & code synthesis
✓Debugging
✓Code editing
✓Repository-aware reasoning
✓Agent & tool-use research
TECHNICAL SPECIFICATION

Overview

M31Genesis (M31-Genesis-AgenticCode-v1) is a native M31 causal language model developed for agentic coding research. It is designed to explore how compact, purpose-built transformer architectures can perform software engineering, code synthesis, editing, and repository-aware reasoning.

Lifecycle Status: Experimental Research Checkpoints

IMPORTANT NOTICE: M31Genesis is explicitly designated as Experimental Research Checkpoints. It is published for research, evaluation, and experimental benchmarking. It is not a stable production product, and must be evaluated rigorously before any operational deployment.

Architecture

The model uses a native M31 decoder-only transformer architecture:

  • —Parameters: 425M-class parameters
  • —Layers: 24 transformer layers
  • —Hidden Size: 1024
  • —Attention Heads: 16 query heads, 8 KV heads (Grouped-Query Attention)
  • —Vocabulary: 65,536 ByteLevel BPE
  • —Positional Encoding: Rotary Position Embeddings (RoPE)
  • —Normalization: RMSNorm
  • —Embedding: Tied word embeddings
  • —Format: Safetensors

Intended Use

The model card specifies the following intended research use cases:

  • —Software engineering and coding
  • —Automated debugging and error analysis
  • —Code editing and diff generation
  • —Repository-aware multi-file reasoning
  • —Agent and tool-use orchestration research

Public Checkpoints & Authoritative Source

Public checkpoints are hosted on Hugging Face at eshanized/M31Genesis. Checkpoints are stored immutably with corresponding configuration, tokenizer, and metadata files.

ROADMAP & MILESTONES

Milestones & active developments

4 / 6 COMPLETED
  • Architecture specification & tokenizer definition
  • Pre-training checkpoint series step-00000500 to latest
  • Hugging Face model card and weights publication
  • Standalone inference and generation runtime (inference.py)
  • Fine-tuning & SFT stage continuation
  • Comprehensive HumanEval and SWE-bench evaluation
CHECKPOINTS & HISTORY

Checkpoint lineage & milestones

Step-wise checkpoints and lineage updates for M31Genesis.

AgenticCode-v1 — Verified Milestone

Experimental checkpoints verified against public repositories.

RELATED RESEARCH

Associated research publications