OpenJarvis: Modular AI & Materials Platform

• OpenJarvis is an open, modular platform defined by agentic computing architectures that enable personalized AI assistance and scalable materials discovery.
• It employs a decomposed, primitive-based design with LLM-guided spec search to optimize performance, reduce latency, and lower operational costs on local devices.
• In materials science, OpenJarvis integrates quantum, classical, and ML workflows to ensure reproducible, high-throughput data-driven discovery.

OpenJarvis refers to a class of open, extensible platforms and architectures for general-purpose agentic computing, with two major contemporary usages: (1) OpenJarvis as a “decomposed” personal AI stack for end-to-end on-device assistance (Saad-Falcon et al., 16 May 2026), and (2) OpenJarvis as a modular, reproducible infrastructure for data-driven materials discovery in computational materials science (Wines et al., 2023, Choudhary, 6 Mar 2025). Both are architecturally characterized by modularity, composability of primitives, and open APIs. The term is also occasionally used to denote unified agentic models in vision-language-action environments (Wang et al., 2024), though this usage is less common in flagship repositories.

1. Compositional Architecture for Personal AI

OpenJarvis, in the context of personal AI, is a decomposed agentic stack designed to deliver state-of-the-art performance in personal assistance while retaining privacy, cost-efficiency, and low-latency on commodity local hardware (Saad-Falcon et al., 16 May 2026). The architecture formalizes the agent stack as a typed specification over five independently swappable primitives:

Primitive	Description	Example Data Type Fields
Intelligence	PLM & decoding params	(model-ID, temperature, quantization, etc.)
Engine	Runtime and execution backend	(backend, batch size, kv cache, etc.)
Agents	Reasoning loop, prompts, tool-calling policy	(agent type, exemplars, turn limits)
Tools/Mem	Integrated utilities, connectors, persistent memory store	(tool set, descriptions, memory backend)
Learning	Optimizer for stack edits, reward weights	(optimizer, reward vector, tolerance)

These primitives are combined in a declarative “spec” $S$, which is end-to-end optimizable. LLM-guided spec search iteratively proposes edits to the spec (as compound or targeted slot-by-slot edits), using a frontier cloud LLM (e.g., Claude Opus 4.6) as a “teacher” to analyze failure clusters and suggest edit actions. Only edits that pass stagnant-gated acceptance on held-out data are retained, ensuring that local deployments maintain or approach the best cloud accuracy, while running exclusively on the device at inference time.

2. Data-Driven Materials Design Infrastructure

In materials science, OpenJarvis is the open variant of the JARVIS (Joint Automated Repository for Various Integrated Simulations) platform, developed under the National Institute of Standards and Technology (NIST), providing reproducible access to quantum, classical, machine learning, and experimental workflows and datasets for materials discovery (Wines et al., 2023, Choudhary, 6 Mar 2025). The architecture organizes resources as:

• Databases: JARVIS-DFT (>80,000 DFT-optimized bulk and 2D materials, phonons, hybrid gaps, and meta-GGA functionals), JARVIS-FF (classical force fields), Wannier tight-binding Hamiltonians, QMC benchmarks, and multimodal experimental data (XRD, STM, neutron diffraction, etc.).
• Electronic Structure/Atomistic Workflows: VASP, QE, DMFT, GW, QMCPACK (QMC), LAMMPS for MD, Wannier90, tight-binding (TB3PY), and quantum computation (VQE, VQD via Qiskit and PennyLane).
• ML Modules: CFID descriptors, ALIGNN and ALIGNN-FF (GNN force fields), AtomVision (CV/segmentation on microscopy), ChemNLP (transformer-based text mining for synthesis/property relationships).
• Open Access and Reproducibility: Released under MIT-style license, all data in JSON/CSV/HDF5, with raw runs and DOIs archived on Figshare/Zenodo, RESTful API endpoints, and Python client.

Code and data access are supported via public repositories (e.g., usnistgov/jarvis and method-documented notebooks), enabling seamless integration into both academic and industrial design pipelines.

3. Optimization and Learning Procedures

For personal AI, OpenJarvis introduces “LLM-guided spec search,” an optimization routine that leverages cloud LLMs to diagnose failures in the deployment spec $S$, cluster them by failure modes, and propose targeted edit actions. Gate-acceptance rules admit an edit only if it offers improvement in the primary cluster and does not regress performance in others by more than a threshold $\epsilon$. The reward for candidate edits is a composite function:

$$R(q,y) = \alpha R_{\mathrm{acc}}(q,y) - \beta \widehat{E}(q,y) - \gamma \widehat{L}(q,y) - \delta \widehat{C}(q,y)$$

where $R_{\mathrm{acc}}$ is accuracy, $\widehat{E}$, $\widehat{L}$, $\widehat{C}$ are normalized energy, latency, and cost, and $(\alpha,\beta,\gamma,\delta)$ are tunable to prioritize between objectives.

In JARVIS, ML surrogates such as ALIGNN are trained via graph neural networks with line-graph representations for directional chemical environments. The loss functions typically used are mean squared error (MSE) or mean absolute error (MAE) with $\ell_2$ regularization. The universal tight-binding Hamiltonian is formalized as:

$S$0

with real-space three-body corrections.

4. Quantitative Performance and Benchmarking

OpenJarvis personal AI stacks close the “local-cloud” accuracy gap present when swapping a local PLM (e.g., Qwen3.5-9B) into existing frameworks. Without stack adaptation, accuracy drops 25–39 pp; OpenJarvis’s decomposed/primitives approach recovers 77%/57% of the loss on PinchBench/GAIA. With LLM-guided spec search, local specs using Qwen3.5-122B achieve an average 80.3% accuracy vs. 83.5% for Claude Opus 4.6 over 8 agentic benchmarks—a 3.2 pp gap—while marginal API cost drops $S$1800$S$2 and latency 4$S$3 (Saad-Falcon et al., 16 May 2026).

In JARVIS, >1,400 benchmarking contributions and >150 unique methods are tracked in a leader-board. The infrastructure supports high-throughput forward/inverse screening, including for superconductors, 2D magnets, defect energetics, and generative atomistic models. JARVIS ensures reproducibility by publishing all workflows, code, data, and benchmarks with rich metadata, DOIs, and continuous integration testing (Wines et al., 2023, Choudhary, 6 Mar 2025).

5. Typical Workflows and APIs

OpenJarvis (personal AI) can be extended or reproduced via:

1. Cloning codebase (github.com/openjarvis/openjarvis), registering models/backends/agents, and specifying specs in TOML.
2. Running LLM-guided spec search as specified via pseudocode, collecting traces, using teacher.diagnose and propose, and evaluating via gates until budget/stagnation.
3. Instrumentation of performance metrics enables Pareto frontier analyses on accuracy, energy, latency, and cost.

JARVIS materials science workflows encompass forward (DFT→property) and inverse (target property→structure) design. Python clients allow querying and downloading data; ML predictors (e.g., ALIGNN) can be installed via pip install jarvis-tools alignn and used to build property surrogates or submit jobs to the benchmarking leader-board.

6. Impact, Limitations, and Future Directions

OpenJarvis for personal AI marks a shift from monolithic, cloud-dependent agentic stacks to end-to-end local optimization, enabling privacy, cost, and latency gains without surrendering model flexibility. However, the gap between local and top cloud model accuracy, while narrowed, is not fully closed; multi-primitives adaptation outperforms single-primitives (e.g., LoRA, prompt evolution) in accuracy and optimization cost. Model registration, engine integration, and agent design remain areas for ongoing extension.

In data-driven materials design, OpenJarvis models a transparent, open-access paradigm for multi-scale, multi-modal science. Experimental data coverage is still more limited than simulation; high-throughput QMC/DMFT is computationally expensive; surrogate models require large training sets for niche properties. Prospective advances include greater integration of quantum computing, more robust generative workflows, and tighter coupling of experiment and computation in real time (Wines et al., 2023, Choudhary, 6 Mar 2025).

7. Related Frameworks and Variants

The term OpenJarvis also appears in unified agentic VLA models, such as OmniJARVIS, which merges self-supervised behavior tokenizers and unified multimodal language modeling for instruction-following agents in open-world environments like Minecraft (Wang et al., 2024). These systems demonstrate that a single compressed tokenization space for language, vision, and action facilitates strong reasoning and sequential execution. Use in such environments reaffirms the shared architectural traits of modularity and unification, although the scale and modality domains differ.

OpenJarvis, across usages, exemplifies the transition toward modular, rigorously benchmarked, and fully open agentic platforms, whether in ML-driven scientific discovery or privacy-preserving assisted computation.