TokenWave AI
TokenWave AI develops AI-powered workflows and research benchmarks for real-world occupations and organizational work. Its research spans benchmarks, synthetic data, agents, reinforcement learning, foundation models, and pretraining-compute allocation, in collaboration with the M-A-P community.
At a Glance
- AI research labs and model developers
- Coding-agent and software-engineering-tool builders
- Organizations evaluating AI employees and workflow automation
- Researchers working on benchmarks, synthetic data, agents, reinforcement learning, and foundation models
AI Tools by TokenWave AI
(1)SWE-CoWork
Coding Agent Benchmark Suite
Discussions
No discussions yet
Be the first to start a discussion about TokenWave AI
Latest News
SWE-CoWork benchmark paper submitted: Benchmarking Coding Agents as Digital Employees in Organizations
SWE-CoWork public repository release reached a release-integrity update and published the benchmark code and evaluation kit
YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality published with Ge Zhang among the authors
xDailyBench: Benchmarking LLMs on Professional Consultation for Real-Life Problems published with Ge Zhang among the authors
Products & Services
An open benchmark for coding agents working inside simulated software organizations. It contains 101 assignments across 101 repositories, 11 programming languages, and 12 application domains; agents recover distributed requirements from colleagues and workplace records, implement changes, verify them, and hand off results.
A benchmark evaluating coding agents as autonomous software test engineers: generating tests from requirements, discovering and localizing coexisting defects, and reporting evidence across 100 instances from 78 Python repositories.
A graduate-level knowledge benchmark covering 285 subfields with sample-weighted and taxonomy-balanced views.
A coding benchmark requiring complete Python libraries to be generated from a requirements document, with correctness determined by the original hidden test suite.
Market Position
TokenWave positions itself as a research organization focused on evaluating AI in realistic, occupation- and workflow-based settings rather than only isolated prompts or issue-resolution tasks. SWE-CoWork differentiates its coding-agent evaluation by placing agents in simulated organizations with distributed requirements, colleagues, workplace records, and end-to-end verification; TestEngineerBench similarly emphasizes requirements-driven exploration and defect localization rather than raw test passing or coverage alone.
Leadership
Founders
Ge Zhang
Identified in Google Scholar as a founder of M-A-P / Tokenwave.AI; the profile describes prior roles at 01.AI and Seed (ByteDance). Ge Zhang is also a co-author on multiple TokenWave-affiliated 2026 research publications, including SWE-CoWork, HarnessDev, StartupBench, xDailyBench, and YuE2.
Executive Team
Ge Zhang
Founder
Google Scholar identifies Ge Zhang as M-A-P / Tokenwave.AI founder and lists prior experience at 01.AI and Seed (ByteDance).
Target Markets
- AI research labs and model developers
- Coding-agent and software-engineering-tool builders
- Organizations evaluating AI employees and workflow automation
- Researchers working on benchmarks, synthetic data, agents, reinforcement learning, and foundation models
- Evaluating coding agents on realistic software-team assignments
- Testing autonomous software testing and defect localization
- Benchmarking knowledge, coding, GUI, agent, and multimodal systems
- Training and evaluating foundation models with synthetic or web-scale instruction data
- Automating professional and organizational workflows
- Research on reinforcement learning and compute-efficient model architectures