Xangel0s/ozyvler-Ozygram-Neural-7B-Coder-MCTS

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Ozygram Neural 7B v2 by Xangel0s is a 7.6 billion parameter neuro-symbolic code intelligence model, built on Qwen2.5-7B-Instruct, with a 32K context length. It specializes in autonomous local code self-healing, bug localization, and transactional integrity through a dual-phase Dream-RSI (MCTS) and Kev Action Engine architecture. This model excels at generating deterministic JSON actions for AST mutative repair, optimized for edge workstations with 8GB VRAM.

Loading preview...

Ozygram Neural 7B v2: Autonomous Code Self-Healing

Ozygram Neural 7B v2 is a specialized 7.6 billion parameter code intelligence model, built upon the Qwen2.5-7B-Instruct base, designed for autonomous local self-healing of code. It integrates a unique neuro-symbolic Monte Carlo Tree Search (MCTS) architecture, enabling advanced bug localization and high-concurrency transactional software engineering.

Key Capabilities & Architecture

This model operates through a dual-phase system:

  • Dream-RSI (Recursive Self-Improvement & MCTS): An internal reasoning chain that explores Abstract Syntax Tree (AST) trajectories, simulates execution, evaluates blast radius across dependent files, detects concurrency deadlocks, and discards destructive paths before committing changes.
  • Kev Action Engine: A deterministic JSON action protocol (ast_mutation, apply_patch) that delivers atomic target/replacement code hunks directly to IDEs or autonomous agents, avoiding conversational prose and fragile line offsets.

Performance & Training

  • Trained via QLoRA on local GPU hardware, achieving a final training loss of 0.005904 and mean token accuracy of >99.4%.
  • Validated on a Master Gym Benchmark Suite with 100% pass rates for AST dependency traversal, concurrency deadlock simulation, and high-throughput boilerplate delegation.
  • Demonstrated successful self-healing on a live local CRM benchmark, autonomously fixing production bugs related to missing prefixes, corrupted regex, date validation, zero-division errors, and worker deadlock issues.

Use Cases

This model is ideal for:

  • Automated bug fixing and code repair in development and production environments.
  • Ensuring transactional integrity and preventing concurrency issues in complex software systems.
  • Integrating autonomous code intelligence directly into IDEs or CI/CD pipelines for self-healing capabilities.
  • Applications requiring deterministic, non-conversational code modifications.

Finetunes
1 models