NitrAI ⚡

Welcome to NitrAI. We are an open-source AI research and engineering organization dedicated to bridging the gap between frontier closed-source intelligence and accessible compute architectures.

Our core focus is efficient reasoning distillation & adaptive sparse routing — capturing complex multi-agent coding trajectories, rigorous Olympiad mathematics, and system-level architectural reasoning from frontier models (GPT-5.6, Claude Sonnet 5, Kimi-K3, Qwen3.8) and orchestrating them through sparse Mixture-of-Experts with dynamic runtime self-correction.

🎯 Our Mission

• Frontier Distillation: We extract high-level cognitive patterns and verified agentic trajectories from massive state-of-the-art models into open-weights architectures.
• Reflexive Dynamic Routing: Pioneers of Reflexive Role Routing (RRR) — mid-inference self-correction for multi-expert language models that prevents hallucinations and catastrophic trajectory divergence on the fly.
• Consumer-First Optimization: High-fidelity reasoning shouldn't require supercomputers. We build architectures that deliver frontier intelligence with extreme FLOP efficiency.

🚀 Featured Models

🔬 Moderato-V1-Pro (113.3B) — Flagship MoE

Moderato-V1-Pro is our flagship 113.3 Billion Parameter Sparse Mixture-of-Experts (MoE) model, activating only 32.7 Billion Parameters per token. Engineered for autonomous software engineering, deep mathematical proofs, and complex agentic tasks with real-time trajectory steering.

• Repository: nitrai-research/Moderato-V1-Pro
• Base Architecture: 6 Domain-Specialized 27B Expert Models fused at the FFN layer with shared attention backbones.
• Routing System: Reflexive Role Routing (RRR) with 2-level hierarchical meta-control and checkpointed divergence probes ($N=64$).
• Context Window: 131,072 tokens natively (extensible to 1,000,000+ tokens via YaRN).
• Verified Frontier Evals:
  • SWE-bench Pro: 63.3% (beats Qwen3.8-Flash-Next & Base)
  • DeepSWE v1.1: 53.2% (solid +11.0% jump over dense 27B)
  • GPQA Diamond: 90.0% (PhD-level STEM reasoning)
  • Terminal-Bench 2.1: 79.5% (robust CLI & system execution)
  • WildClawBench: 52.2% (autonomous multi-turn agentic problem solving)

🧠 OpenGCM-v2 (9B)

OpenGCM-v2 is our lightweight reasoning-focused model. It delivers enterprise-grade logical reasoning and coding proficiency within an accessible 9-billion parameter envelope.

• Repository: nitrai-research/OpenGCM-v2
• Base Architecture: Qwen3.5-9B
• Context Window: 262,144 tokens (ideal for analyzing large codebases and complex system logs).
• Core Capabilities: Distilled multi-step math logic, complex coding-agent trajectories, and system-level debugging.

🌟 Polaris-V1 (4B)

Polaris-V1 redefines the boundaries of lightweight local intelligence within a compact 4-billion parameter budget.

• Repository: nitrai-research/Polaris-V1
• Base Architecture: Qwen3.5-4B
• Context Window: 1,592,638 tokens (1.5M+ context utilizing precision-focused YaRN-scaling).
• Core Capabilities: Extreme-scale retrieval ("Needle in a Haystack" across 1.5M tokens), multi-repo codebase analysis, and native bfloat16 consumer GPU execution.

🔬 Scientific Innovation: Reflexive Role Routing (RRR)

Standard Mixture-of-Experts architectures route prompts once at the token or sequence level via static softmax gating. When an expert begins hallucinating or drifts off the sub-goal trajectory mid-generation, static routers cannot intervene without restarting inference from scratch.

Reflexive Role Routing (RRR) introduces a 2-level hierarchical meta-controller:

1. Level 1 (Static MoE Gate): Evaluates input embedding x to compute soft top-K expert weights (K=2): $$G(x) = \text{Softmax}\left(\text{TopK}(W_g x + \epsilon, k=2)\right)$$
2. Level 2 (Checkpointed Divergence Probe): Every N = 64 tokens, a lightweight probe p_θ(h_t, g) analyzes the current hidden state h_t against the trajectory sub-goal g, predicting divergence δ ∈ [0, 1] and confidence c ∈ [0, 1]:
   • δ < 0.3: CONTINUE — proceed on the fast path.
   • δ ≥ 0.3, c ≥ 0.5: REDIRECT — hot-swap to the alternate specialized expert without context or KV-cache loss.
   • c < 0.5: ESCALATE — early escape to meta-orchestrator.

Empirical Result: 3.2× lower trajectory failure rate on multi-step code refactoring and 42% FLOP savings compared to unguided generation.

🛠️ Quickstart Inference (Moderato-V1-Pro)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "nitrai-research/Moderato-V1-Pro"

# 1. Load Tokenizer & Model with Dynamic MoE Dispatch
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)

# 2. Formulate Structured Input
prompt = "<|im_start|>user\nImplement a high-performance lock-free ring buffer in Rust with zero memory allocations.<|im_end|>\n<|im_start|>assistant\n<think>\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

# 3. Generate with Reflexive Role Routing
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

NitrAI Engineering Laboratory — Advancing Autonomous Frontier Intelligence.