NitrAI captures multi-step coding trajectories and mathematical proofs from frontier models—packaging them into efficient 9B & 4B architectures designed for local execution on everyday GPUs and CPUs.
ollama run nitrai/opengcm-v2:9b
Our work focuses on extracting structural cognitive trajectories into efficient open architectures without datacenter hardware dependencies.
We extract multi-step reasoning traces, coding trajectories, and mathematical proof structures from large frontier models (GPT-5.5, Claude-Fable-5, GLM-5.2) into open-weights base architectures.
Models are quantized and compiled with custom CUDA and Apple Metal kernels to ensure low-latency inference on desktop and laptop hardware.
Architectures are specifically fine-tuned for agentic multi-turn workflows, long-context code comprehension, and formal Lean 4 proof validation.
Lightweight open models designed for local deployment and specialized reasoning tasks.
OpenGCM-v2 is our flagship reasoning model, engineered to deliver structured code synthesis and formal logical proofs within a 9-billion parameter budget.
# Run via Ollama (Q4_K_M)
ollama run nitrai/opengcm-v2
# Python SDK
from nitrai import NitrClient
client = NitrClient(model="opengcm-v2-9b")
res = client.reason(prompt="Analyze FFI memory layout")
print(res.text)A four-stage engineering pipeline for capturing reasoning patterns and compiling open weights.
Extracting multi-agent tool execution, code edits, and proof trees from frontier models.
Automated filtering using Lean 4 theorem provers and sandbox test suites.
Attention interpolation scaling effective context to 1.59M tokens.
Compiling bfloat16 weights into native Metal/CUDA kernels for local hardware.
We record synthetic reasoning traces from large closed models. NitrAI captures internal thinking blocks, self-correction loops, and multi-file code investigation paths rather than simple prompt-completion pairs.
Calculate expected memory consumption and inference speeds based on target hardware.
Select a task to inspect distilled reasoning streams and code output.
Compatible with standard local inference engines and frameworks.
# Pull and run OpenGCM-v2 (9B) locally via Ollama
ollama run nitrai/opengcm-v2
# Run query with reasoning trajectory enabled
ollama run nitrai/opengcm-v2 "Analyze memory leaks in this Rust binary..."Test 1.59M token context windows and 2026 frontier distillation on local GPU hardware.