Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Materials science agent MatOS goes open source with 3,135 datasets and an 80.8% edge over Codex CLI on materials tasks

MatOS, a research agent built by the AI materials institute at Beijing's Zhongguancun institutions, is now open source with code on GitHub and a technical report on arXiv, backed by 3,135 datasets, 386 skills and 48 materials programs. With GPT-5.4 as the shared base model it beat Codex CLI by 80.8% on materials tasks, while its cheaper configuration averaged about $0.85 per task.

Published

MatOS, a research agent built for materials discovery, has been released as open source, with code on GitHub and a technical report on arXiv. The system was built by the AI materials institute at Beijing Zhongguancun College and the Zhongguancun Institute of Artificial Intelligence, which first presented it in their annual research progress in September.

The core idea is to register data, knowledge, tools, skills, roles and workflows as a unified set of scientific assets. According to QbitAI, the library now holds 3,135 datasets, 386 skills, 104 workflows and 51 expert roles, and integrates 48 materials computation programs.

Each asset carries a unique identifier, declared dependencies and provenance, and can either store content directly or reference an external database or simulation package. Assets are linked by typed relations such as dependency, invocation, evidence and verification, so the agent assembles the capabilities a task needs instead of rebuilding a pipeline every time a new database or tool appears.

At execution time, MatOS runs a study through a stateful execution graph. Every node states its goal, inputs and outputs, completion conditions and the assets it uses; the system schedules and checks along the graph and hands only genuinely judgemental steps — method selection, plan revision, anomaly explanation — to the language model. Steps that need a researcher's confirmation can be set as human nodes.

The evaluation offers quantified comparisons. On ResearchClawBench, an end-to-end autonomous research benchmark, MatOS scored 19.46 on average with GPT-5.4 as the shared base model, 11.15% above the best baseline MIRA and 33.6% above Codex CLI; on materials tasks the gap to Codex CLI widened to 80.8%. Swapping in the cheaper DeepSeek-V4-Flash, MatOS averaged 19.20 and still beat every baseline.

On MatTools, a benchmark for using materials software, MatOS reached a 97.96% full-question success rate with DeepSeek-V4-Pro as the common base model, against 95.92% for Codex CLI and 89.80% for Claude Code. On SciAgentGym, a multi-step tool-use benchmark, its case success rate was 59.5% versus 57.1% for the other two.

The case studies span solid electrolytes and semiconductor packaging alloys. For electrolytes, MatOS used a layered screening framework — a composition-only model for coarse filtering, a graph neural network for structure-based property prediction, then molecular dynamics for ion transport — and narrowed 30,364,908 initial structures to 97 candidates, with a top computed conductivity of 59.0 mS/cm. In an alloy formulation search for semiconductor packaging aimed at melting near 300 degrees Celsius, the best score across 286 candidates rose from 0.534 for the starting formula to 0.927, with the top Pb-Ag-Pt-Sn alloy predicted to start melting at 297.75 degrees Celsius and reach an 80% liquid fraction after a 4.55 degree rise.

Cost is part of the pitch: a DeepSeek-V4-Flash configuration averaged about $0.85 per task, less than a third of Codex CLI's $2.70 while scoring higher, while the GPT-5.4 configuration took the highest score at about $3.78. Together the two configurations extend the benchmark's cost-score Pareto frontier.

The mechanism worth watching is asset evolution. User corrections, failed tool calls, repeated operations and workflow results can all trigger a revision, and a proposed change must pass structural checks and a model-based semantic review before adoption. Because the experience lives in assets outside the model while the base model's parameters stay fixed, one MatOS stack has been paired with both GPT-5.4 and DeepSeek models. That makes research capability something that can be versioned, reused and audited like a codebase rather than tied to one generation of foundation model.

Why it matters

Turning research workflows into versioned, reviewable assets rather than model weights lets materials agents carry over across foundation models and lowers the unit cost of long-running scientific tasks.

MatOSAI AgentOpen SourceMaterials Science
Back to realtime news

Nearby Updates

All

11/17, 15:43

YC-Backed Scalar Field Logs $60M in Trading Activity in Four Months of AI Agents

YC-backed startup Scalar Field reached $60 million in trading activity in the first four months of its AI agent rollout, according to CXOToday. The figure puts an agent product in front of real money rather than demos, though the report does not break out user counts, asset types or revenue.

11/17, 16:01

Australian creatives launch Not AI's to Take campaign to defend copyright

Australian arts, screen, media and music organizations have launched the Not AI's to Take campaign, pressing the federal government to defend copyright against unlicensed use of creative work in AI training. Backers include the Australian Society of Authors and the Australian Publishers Association, and the push is aimed at countering AI industry lobbying for looser rules.

11/17, 17:52

Appian ships an AI agent control plane and Ask Appian agentic interface

Appian used its annual Appian Europe conference in London on November 17 to roll out agent-management capabilities that track, orchestrate and optimise AI agents inside deterministic business processes, alongside a natural-language interface called Ask Appian. The company cites a 2026 study with Harvard Business Review Analytics Services finding that only 16% of respondents report a high degree of measurable value from AI.

11/17, 09:28

DeepSeek rations free chat over compute shortage as V4.1 Flash tops OpenRouter at 61.2T tokens

DeepSeek is rationing its free app and website, dropping long-idle conversations and capping single inputs at about 40,000 characters because of insufficient compute. The limits land as its V4.1 Flash 1017 model leads OpenRouter with 61.2 trillion tokens in a week, though DeepSeek's own endpoint serves only 12.4% of that traffic.