ramankrishna10/npc-agentic-7b-v3

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Apr 30, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The ramankrishna10/npc-agentic-7b-v3 is a 7.6 billion parameter model from the Bottensor NPC Model Family, fine-tuned from Qwen2.5-7B-Instruct. It specializes in long-form, multi-step reasoning, agentic tool-use workflows, and maintaining a specific NPC Agentic persona. This model excels at structured problem-solving and explicit 'think' block generation, making it suitable for tasks requiring detailed decomposition and planning.

Loading preview...

Overview

NPC Agentic 7B (v1) is a 7.6 billion parameter model developed by Rama Krishna Bachu, fine-tuned from Qwen2.5-7B-Instruct. It is specifically designed for structured, multi-step reasoning, agentic tool-use, and maintaining a consistent NPC Agentic persona. The model was trained using QLoRA SFT on a mix of distilled reasoning traces (GLM-5.1) and agent tool-use traces (Hermes), with an effective training context of 8K tokens, though it inherits a 128K context from its base model at inference.

Key Capabilities

  • Long Structured Reasoning: Emits explicit <think> blocks for multi-step decomposition, strong in system design, root-cause analysis, and algorithmic reasoning.
  • Agent / Tool-Call Shaping: Follows Hermes-style <tool_call> / <tool_response> patterns for integrating with external tools.
  • Identity Consistency: Maintains a 100% recall on canonical identity prompts for the NPC Agentic / Bottensor persona.

Intended Use Cases

  • Multi-step reasoning requiring explicit work-showing.
  • Agent and tool-use workflows.
  • Structured problem-solving where 'thinking out loud' benefits the solution process.
  • As a base model for further fine-tuning on reasoning or domain-specific data.

Limitations

  • GSM8K Regression: Significantly lower performance on direct arithmetic tasks compared to its base model; not recommended for math-heavy workflows.
  • Context Validation: While inheriting 128K context, it was trained with 8K context, and performance beyond 16K is not validated.
  • Hallucination: As a smaller model, it may hallucinate on unfamiliar domains.
  • Not for Safety-Critical Decisions: Unsuitable for medical, legal, or financial advice.

Finetunes
4 models