mira-agent-instinct

Mira AI Agent Model – Instinctive AGI / RSI

Mira is an agent-oriented LLM with instinctive reasoning compression and safe recursive self-improvement loops. Designed for long-horizon tasks, tool use, and self-play fine-tuning with eval gating.

Overview

• Instinct mode: fast heuristic policy that reduces reasoning tokens while preserving first-attempt accuracy.
• Agent layer: smolagents / Transformers Agent Toolkit compatible.
• RSI: offline self-play fine-tuning with DPO, frozen improvement operator, lineage tracking.
• Safety: gated repo, guardrails, audit log, explicit limitations.

Model Details

• Base model: Qwen/Qwen3-8B
• Architecture: decoder-only transformer
• Context length: 32k
• Training: supervised fine-tune on agent trajectories + self-play synthetic data
• Quantized variants: Q4_K_M, Q8_0

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "slab1/mira-agent-instinct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [
    {"role": "user", "content": "Plan a 3-step research workflow and execute step 1"}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")
outputs = model.generate(**inputs)

smolagents

from smolagents import CodeAgent, HfApiModel

model = HfApiModel(model_id="slab1/mira-agent-instinct")
agent = CodeAgent(model=model, tools=[...])
agent.run("...")

Demo

See the HuggingFace Space: slab1/mira-agent-instinct-demo

Evaluation

Results in .eval_results/: * Terminal-Bench 2.1 * SWE-Bench Verified * AgentBench

Limitations

• Not AGI. Agent-oriented LLM with heuristic compression.
• Recursive self-improvement is offline and eval-gated.
• Tool use requires guardrails.

Citation

@misc{mira-agent-instinct,
  title={Mira AI Agent Model – Instinctive AGI / RSI},
  author={slab1},
  year={2026}
}

License

Apache-2.0