TECHNOLOGY

Models that understand how code runs on your hardware.

Models read code as text; they do not know how it runs on real hardware. We run the same code across architectures, devices and inputs and learn its performance, resource use and execution paths: a computational world model.

COMPUTATIONAL WORLD MODEL

From code to how it behaves on hardware.

  1. 01

    Code in

    The same code, across scenarios and inputs.

  2. 02

    Run and observe

    Real runs on CPUs, GPUs, domestic Chinese chips and other accelerators, with their behavior recorded.

  3. 03

    Learn the world model

    Model the code’s performance, memory and resources, and execution paths on each kind of hardware.

Meta’s Code World Model ↗ simulates what code computes (variables and state) in one Python environment. Ours models how the same code runs across chips and devices: speed, memory and execution paths.

RESEARCH

Three lines of research, each with papers and results.

Vulnerability detection

We model programs from both what they do at run time and what the code says. That found more than 100 vulnerabilities in 20+ open-source projects and 34 new CVEs.

ICSE 2024 · ISMM 2025 · IEEE TIFS 2020

Verifiable coding agents

Program analysis keeps the agent in check, so its conclusions can be verified. On 72 binary-analysis tasks, BinReX succeeds nearly twice as often as OpenAI Codex with IDA. EXRAY finds 77.9% of known vulnerabilities, and found 24 previously unknown RCE vulnerabilities in 130 applications.

USENIX Security 2026 · IEEE S&P 2027

BENCHMARKS

Measured against the alternatives.

Binary analysis

1.9×success rate, against OpenAI Codex with IDA

Task success rate

BinReX83.3%
Codex + IDA43.1%

Total analysis time (lower is better)

BinReX36.1 h
Codex + IDA87.9 h
72 tasks across 7 kinds of scenario. USENIX Security 2026 ↗

Training memory

6.5×larger model on one GPU

Largest model on one 32 GB V100

Ours39.5B
Mainstream offloading6B
In a customer’s production environment, with no change to training code. Case study: A leading e-commerce and cloud company →

Training throughput

4.5×training throughput (best case)

Against offloading-only methods (best case)

Ours4.5×
Offloading only1×
Training semantics unchanged. IEEE TPDS 2026 ↗

Reinforcement-learning autotuning

>11×faster Halide image pipelines (some benchmarks)

Halide image pipelines (some benchmarks)

Ours>11×
Default scheduler1×

TVM neural-network code

Ours1.74×
Default schedule1×

LLVM code-size reduction

Ours1.57×
-Oz1×
Each against its default; the LLVM result is the best on 90% of benchmarks. CC 2022 ↗

Big/little-core scheduling

−63.5%energy

Energy (lower is better)

Ours−63.5%
Default Linux schedulerBaseline

Load time (lower is better)

Ours−45%
Default Linux schedulerBaseline
Measured on 500 popular web pages. Case study: A global chip-design company →

Vulnerability-detection false positives

About −50%false-positive rate

False-positive rate (lower is better)

OursAbout −50%
Previous toolBaseline
Running in a customer’s commit and merge flow. Case study: A telecom operator’s cloud →

WHERE IT IS USED

Where it is used

Vulnerability detection

Yi Guard checks an agent against your policy before it calls a tool, runs code or touches a file, and stops dangerous actions.

Explore Yi Guard

Optimization and migration

We deploy AI products to the compute centre you choose, move them onto its chips and tune them; Yi Code refactors and rewrites code in a sandbox.

Explore Yi Code

Verifiable coding agents

Yi Eval tests an agent before launch and after every change, and every finding comes with evidence you can check.

Explore Yi Eval

BLOG

From the blog

All posts

PAPERS

Papers

Next step

AI that understands how code runs.

Contact