How we’re unifying AI frameworks and laying the groundwork for privacy-preserving machine learning

In the rapidly evolving landscape of artificial intelligence and Web3, one challenge remains paramount: how do we create AI systems that are both powerful and trustworthy? At Cortex, we’ve spent the past six months building the foundational infrastructure to solve this exact problem. Our work bridges the gap between cutting-edge AI frameworks and privacy-preserving technologies, creating a platform where machine learning models can be both high-performing and verifiable.

Half-Year Achievements: Laying the Foundation

Expanding Our Privacy Technology Horizon

The privacy technology space moves at lightning speed, and staying current requires constant research and evaluation. We’ve been deeply immersed in exploring the latest Zero-Knowledge Proof (ZKP), ZK Machine Learning (ZKML), and Trusted Execution Environment (TEE) solutions, including hands-on work with EZKL, Taiko, and ZKSync. This isn’t just theoretical research — we’ve been actively integrating and optimizing ZKRollup code to ensure our stack leverages the most efficient and advanced privacy-preserving technologies available.

Unifying AI Frameworks: TVM and PyTorch Integration

A major breakthrough this half-year has been our expansion of Model Representation Toolkit (MRT) frontend support. AI developers work across multiple frameworks, and we believe they shouldn’t be limited by interoperability constraints.

TVM Frontend Support

• Achieved compatibility between Relay/Relax API and MRT Symbol
• Successfully tested with CNN models including ResNet and YOLO
• Created a seamless pathway from TVM to our verifiable inference pipeline

PyTorch Frontend Support

• Implemented compatibility between ExportedProgram API and MRT Symbol
• Validated with the same rigorous CNN model testing
• Now supporting one of the most popular ML frameworks in the ecosystem

Runtime Revolution: Flexibility and Determinism

We completely refactored the MRT runtime code to support dynamic switching between PyTorch and TVM frontends. This isn’t just a quality-of-life improvement — it enables deterministic model inference, a critical requirement for verifiable AI systems where computation must be reproducible and consistent across different environments.

Beyond the major features, we’ve dedicated significant effort to polishing the developer experience:

• Enhanced logging for better debugging and monitoring
• Code refactoring for maintainability and performance
• API adjustments based on real-world usage feedback
• Model-specific issue resolution

Looking Ahead: Our Strategic Roadmap

Near-Term Objectives

Our immediate focus areas build directly on the foundations we’ve established:

Continuous Technology Evaluation
We maintain our commitment to staying at the forefront of privacy technology, continuously assessing new ZKP, ZKML, and TEE solutions as they emerge.

Large Language Model Support
The AI landscape is shifting toward LLMs, and so are we. We’re adding frontend components for transformers and llama.cpp, bringing verifiable AI to the large language model domain.

Core Infrastructure Completion
We’re finalizing the MRT Symbol structure with comprehensive Symbol, Function, and Graph constructs, creating a robust foundation for complex model representations.

Hardware Performance Optimization
Updating CVMRuntime to support the latest CUDA architecture ensures our solutions deliver maximum performance on modern GPU hardware.

2026: The Year of Scale and Performance

Q1-Q2 2026: Architectural Revolution

• Comprehensive refactor of MRT lower architecture
• Support for at least one LLM frontend framework
• Foundation for enterprise-scale model deployment

Q3-Q4 2026: Hardware Evolution

• CVMRuntime library overhaul
• Upgrade to latest CPU/GPU architecture
• Performance optimization for next-generation hardware

Why This Matters

The convergence of AI and blockchain isn’t just a technological curiosity — it’s becoming a necessity. As AI systems make increasingly important decisions, the ability to verify their computations becomes critical. Whether it’s ensuring a model hasn’t been tampered with, proving that inference was performed correctly, or maintaining privacy while leveraging collective intelligence, verifiable AI addresses fundamental trust challenges in our increasingly automated world.

About Cortex 😇

Cortex’s main mission is to provide state-of-the-art machine-learning models on the blockchain in which users can infer using smart contracts on the Cortex blockchain. One of Cortex’s goals also includes implementing a machine-learning platform that allows users to post tasks on the platform, and submit AI DApps (Artificial Intelligence Decentralized Applications).

Cortex is the only public blockchain that allows the execution of nontrivial AI algorithms on the blockchain. MainNet has launched. Go build!

TestNet

| Block Explorer — Cerebro| Mining Pool | Remix Editor | Software | ZkMatrix

Social Media

| Website | GitHub | Twitter | Facebook | Reddit | Kakao | Mail | Discord

Telegram Groups

| English | Korean | Chinese | Russian|Turkish

Cortex’s Half-Year Progress and Roadmap was originally published in Cortex Labs on Medium, where people are continuing the conversation by highlighting and responding to this story.

How we’re unifying AI frameworks and laying the groundwork for privacy-preserving machine learning

In the rapidly evolving landscape of artificial intelligence and Web3, one challenge remains paramount: how do we create AI systems that are both powerful and trustworthy? At Cortex, we’ve spent the past six months building the foundational infrastructure to solve this exact problem. Our work bridges the gap between cutting-edge AI frameworks and privacy-preserving technologies, creating a platform where machine learning models can be both high-performing and verifiable.

Half-Year Achievements: Laying the Foundation

Expanding Our Privacy Technology Horizon

The privacy technology space moves at lightning speed, and staying current requires constant research and evaluation. We’ve been deeply immersed in exploring the latest Zero-Knowledge Proof (ZKP), ZK Machine Learning (ZKML), and Trusted Execution Environment (TEE) solutions, including hands-on work with EZKL, Taiko, and ZKSync. This isn’t just theoretical research — we’ve been actively integrating and optimizing ZKRollup code to ensure our stack leverages the most efficient and advanced privacy-preserving technologies available.

Unifying AI Frameworks: TVM and PyTorch Integration

A major breakthrough this half-year has been our expansion of Model Representation Toolkit (MRT) frontend support. AI developers work across multiple frameworks, and we believe they shouldn’t be limited by interoperability constraints.

TVM Frontend Support

• Achieved compatibility between Relay/Relax API and MRT Symbol
• Successfully tested with CNN models including ResNet and YOLO
• Created a seamless pathway from TVM to our verifiable inference pipeline

PyTorch Frontend Support

• Implemented compatibility between ExportedProgram API and MRT Symbol
• Validated with the same rigorous CNN model testing
• Now supporting one of the most popular ML frameworks in the ecosystem

Runtime Revolution: Flexibility and Determinism

We completely refactored the MRT runtime code to support dynamic switching between PyTorch and TVM frontends. This isn’t just a quality-of-life improvement — it enables deterministic model inference, a critical requirement for verifiable AI systems where computation must be reproducible and consistent across different environments.

Beyond the major features, we’ve dedicated significant effort to polishing the developer experience:

• Enhanced logging for better debugging and monitoring
• Code refactoring for maintainability and performance
• API adjustments based on real-world usage feedback
• Model-specific issue resolution

Looking Ahead: Our Strategic Roadmap

Near-Term Objectives

Our immediate focus areas build directly on the foundations we’ve established:

Continuous Technology Evaluation
We maintain our commitment to staying at the forefront of privacy technology, continuously assessing new ZKP, ZKML, and TEE solutions as they emerge.

Large Language Model Support
The AI landscape is shifting toward LLMs, and so are we. We’re adding frontend components for transformers and llama.cpp, bringing verifiable AI to the large language model domain.

Core Infrastructure Completion
We’re finalizing the MRT Symbol structure with comprehensive Symbol, Function, and Graph constructs, creating a robust foundation for complex model representations.

Hardware Performance Optimization
Updating CVMRuntime to support the latest CUDA architecture ensures our solutions deliver maximum performance on modern GPU hardware.

2026: The Year of Scale and Performance

Q1-Q2 2026: Architectural Revolution

• Comprehensive refactor of MRT lower architecture
• Support for at least one LLM frontend framework
• Foundation for enterprise-scale model deployment

Q3-Q4 2026: Hardware Evolution

• CVMRuntime library overhaul
• Upgrade to latest CPU/GPU architecture
• Performance optimization for next-generation hardware

Why This Matters

The convergence of AI and blockchain isn’t just a technological curiosity — it’s becoming a necessity. As AI systems make increasingly important decisions, the ability to verify their computations becomes critical. Whether it’s ensuring a model hasn’t been tampered with, proving that inference was performed correctly, or maintaining privacy while leveraging collective intelligence, verifiable AI addresses fundamental trust challenges in our increasingly automated world.

About Cortex 😇

Cortex’s main mission is to provide state-of-the-art machine-learning models on the blockchain in which users can infer using smart contracts on the Cortex blockchain. One of Cortex’s goals also includes implementing a machine-learning platform that allows users to post tasks on the platform, and submit AI DApps (Artificial Intelligence Decentralized Applications).

Cortex is the only public blockchain that allows the execution of nontrivial AI algorithms on the blockchain. MainNet has launched. Go build!

TestNet

| Block Explorer — Cerebro| Mining Pool | Remix Editor | Software | ZkMatrix

Social Media

| Website | GitHub | Twitter | Facebook | Reddit | Kakao | Mail | Discord

Telegram Groups

| English | Korean | Chinese | Russian|Turkish

Cortex’s Half-Year Progress and Roadmap was originally published in Cortex Labs on Medium, where people are continuing the conversation by highlighting and responding to this story.

MRT/CVM

1. 调整Trace API接口，将transform pass封装，增加 fuse,calibrate,quantize 和 export 方法;

2. Trace API 增加 dataset，statistic 绑定，增加 accuracy eval 接口;

3. circom 适配 alexnet，dense，修复 padding shape，concat 算子等问题;

4. 修复 dropout + tuplegetitem tvm 运行出错，修复 use too much shared data 显存错误;

5. ZKML 支持 shufflenet, densenet, vgg, squeezenet，和原始浮点模型精度误差在1%~2%;

ZkRollup

1. 实现VM 里面memory和storage 部分的数据读取和修改;

2. 根据不同的Operand类型(寄存器，stack，代码区等)进行不同的数据处理;

3. 实现VM 的一个cycle的基本处理流程：取指， 译码，执行;

4. 实现基本的算术指令的具体逻辑（加减乘除），二元运算，移位运算，vm 上下文操作和指针运算;

Cortex 全节点

1. cmd/evm: fix some issues with the evm run command (#28109);

2. trie: add getter for preimage store in trie.Database (#28155);

3. cmd/devp2p: use bootnodes as crawl input (#28139);

4. cmd/geth, internal/flags: print envvar config source and bad names (#28119);

5. metrics: add support for enabling metrics from env vars (#28118);

市场运营

1.协助MRT进行测试；

2.协助CVM进行测试；

3.NFT 艺术家合作对接；

4.持续搜集、翻译、原创深度相关内容，NFT及DeFi等领域，尝试开拓全新的社区、媒体内容运营方式；

5.持续对市场热门项目进行关注并分析项目逻辑及经济模型等。

| Website | GitHub | Twitter | Facebook | Reddit | Kakao | Mail | Discord

【周报】Cortex团队技术周报20230926 was originally published in Cortex Labs on Medium, where people are continuing the conversation by highlighting and responding to this story.

MRT/CVM

1. 重构完成 quantize sampling 代码，原浮点精度为 86.25% ，量化后为 80.62%;

2. 重构后代码 quantize rules 采用注册机制，修复此前 add/sub scale 不统一的情况;

3. 添加 Python numpy 格式下的推理与 FuseConstant pass;

4. FuseConstant 采用 tvm runtime infer实现，移除mxnet, numpy依赖;

5. 增加 FuseNaiveMathmatic pass, 消除 加0，乘1等情况;

6. ZKML 新增 mobilenet, alexnet 等模型支持;

ZkRollup

1. 定义opcode需要满足的属性和一些基本Opcode（ADD,MUL等），实现对应从U256 解码，并encode到u64和bytes8;

2. 确定操作数的可能类型，生成了一个全局的操作码解码查找表;

Cortex 全节点

1. core/forkid: skip genesis forks by time;

2. simplify storage trie update and commit;

3. informational metrics (prometheus, influxdb, opentsb);

4. account/abi: handle solidity panic revert (#27868);

市场运营

1.对ZKML产品进行使用测试；

2.调研跨链桥的生态和进展；

3.NFT 艺术家合作对接；

4.持续搜集、翻译、原创深度相关内容，NFT及DeFi等领域，尝试开拓全新的社区、媒体内容运营方式；

5.持续对市场热门项目进行关注并分析项目逻辑及经济模型等。

| Website | GitHub | Twitter | Facebook | Reddit | Kakao | Mail | Discord

【周报】Cortex团队技术周报20230912 was originally published in Cortex Labs on Medium, where people are continuing the conversation by highlighting and responding to this story.

MRT/CVM

1. ZKML 调优 ResNet 模型精度，提升 14% 准确率

2. ZKML 调整 Trace API，增加数据集绑定，预处理，后处理信息等接口

3. ZKML 完成数据集 ImageNet，coco 等接口，并下载更多数据集

4. ZKML 调研 ezkl，跑通流程并对比 ZKML Circom benchmark 性能

ZkRollup

1. 实现zkcvm的opcode中

2. 调研并跑通新的 poseidon hash 实现以及 poseidon-hash circuit in halo2

3. 调研跟进最新zkevm开发(cariovm,fuelvm,syncvm,midenvm)

4. 实现 zkcvm 的 OPCode

5. 尝试实现 zkcvm 的execution layer（修改共识，出块，smt等部分）

Cortex 全节点

1. core/rawdb: allocate database keys with explicit size to avoid slice

2. core: ensure txindex will be triggered at least once (#27847)

3. ignore genesis block on importChain

4. trie: reduce allocs in recHash (#27770)

5. all: update golang/x/ext and fix slice sorting fallout (#27909)

市场运营

1.整理当前ZKML的项目调研报告；

2.分析项目报告内容整理适合Cortex的产品；

3.调研跨链桥的生态和进展；

4.NFT 艺术家合作对接；

5.持续搜集、翻译、原创深度相关内容，NFT及DeFi等领域，尝试开拓全新的社区、媒体内容运营方式；

6.持续对市场热门项目进行关注并分析项目逻辑及经济模型等。

| Website | GitHub | Twitter | Facebook | Reddit | Kakao | Mail | Discord

【周报】Cortex团队技术周报20230829 was originally published in Cortex Labs on Medium, where people are continuing the conversation by highlighting and responding to this story.

MRT/CVM

1. ZKML 训练 RNN 语言序列预测模型;

2. ZKML subgraph 接口存储实现优化，父子trace共用节点指针（hash）;

3. ZKML 调研精度参数，并测试参数对量化模型的影响;

4. ZKML Make config class for quantization procedure(with statement);

ZkRollup

1. 测试运行了boojum的sha256 电路，preimage < 2kb 时候成功（很快）,4kb 和8kb会卡住;

2. L1Client 分成quering client 和signing client和L1交互，fix重构导致的其他模块的错误;

3. add the generic support for RocksDB storage;

4. Some chore refactor:migrate workspace dependencies, use basic_type::web3 as a replacement for direct ethers-rs;

5. 电路benchmark 结果很差（realease 模式）,需要考虑切换到boojum (fri) 或者halo2;

Cortex 全节点

1. core/state: move UpdateContractCode before the trie hash is computedore/state: move UpdateContractCode before the trie hash is computed;

2. p2p: move ping handling into pingLoop goroutine (#27887);

3. all: remove trailing whitespace (#27741);

4. build: upgrade -dlgo version to Go 1.20.7 (#27835);

5. cmd/bootnode: fix timer leak (#27754);

市场运营

1.调研当前ZKML的生态和进展；

2.调研当前MPC钱包的技术进展；

3.调研跨链桥的生态和进展；

4.NFT 艺术家合作对接；

5.持续搜集、翻译、原创深度相关内容，NFT及DeFi等领域，尝试开拓全新的社区、媒体内容运营方式；

6.持续对市场热门项目进行关注并分析项目逻辑及经济模型等。

| Website | GitHub | Twitter | Facebook | Reddit | Kakao | Mail | Discord

【周报】Cortex团队技术周报20230815 was originally published in Cortex Labs on Medium, where people are continuing the conversation by highlighting and responding to this story.

MRT/CVM

1. 完成 coco，voc，cifar10 数据集加载的功能;

2. circom生成器RightShift和MulScalar参数对齐。类型注解，Symbol类型混淆的问题修复;

3. 生成图优化，拓扑排序后丢失无效节点。scalar整合到op内作为属性，原参数仍输出到input.json，fix circom编译修复;

4. 增加NLP模型训练，长序列数值预测，转换定点化模型并调试问题;

ZkRollup

1. 完成zkx-contract 重构成 diamonds-proxy 架构，并且更新部署脚本和工具脚本;

2. 调整 L1 节点和 postgre docker 启动（方便script 启动关闭）;

3. Golang实现zkx-cli infrastructure 工具：用于development 更加方便的启动;

4. 更新 contract 和 config crates 适配新的合约;

5. 修复tx请求发送过多，Nonce 不匹配的问题;

6. 整体重构zkx-contract 从原本架构（普通proxy + 合约分割（包含addtional）） 到EIP-2535;

7. 调研 zksync-era 新系统 PLONK + FRI：boojum;

Cortex 全节点

1. cmd/devp2p: atomic types used (#27755);

2. core: replace instances of ‘suicide’ with ‘selfdestruct’ to improve code consistency;

3. metrics: NilTimer should still run the function to be timed (#27723);

4. cmd/bootnode, p2p: support for alternate mapped ports (#26359);

市场运营

1.调研当前ZKML的生态和进展；

2.调研当前MPC钱包的技术进展；

3.调研跨链桥的生态和进展；

4.NFT 艺术家合作对接；

5.持续搜集、翻译、原创深度相关内容，NFT及DeFi等领域，尝试开拓全新的社区、媒体内容运营方式；

6.持续对市场热门项目进行关注并分析项目逻辑及经济模型等。

| Website | GitHub | Twitter | Facebook | Reddit | Kakao | Mail | Discord

【周报】Cortex团队技术周报20230801 was originally published in Cortex Labs on Medium, where people are continuing the conversation by highlighting and responding to this story.

ML&MRT&CVM 研发

1. 设计和测试模型量化过程中的logs接口位置和参数
2. 修复zkml 里的bugs，并进行代码调整和优化
3. 实现image input_data单独进行scale，round的接口
4. 模型转换成circom后，修改input.json并设置input value
4. resnet18 resnet34 resnet50 resnet101 resnet152转换格式成功

ZkRollup

1. 实现和zkx netwrok 交互的js sdk: zkx.js.
2. 使用handlebars 重新render single verifier 和aggregate verifier 的solidity代码
3. 添加health check api用于监测数据库和web3 连接的可用性
4. 修改price fetcher 从第三方api(coingecko, coinmarketcap…)获取数据的流程，添加validator验证token是否可被用于zkx network 交易的fee使用

Cortex 全节点

1. crypto/secp256k1: define NDEBUG only if not defined #1600
2. update leveldb metrics collection and track state changes in state db
3. database put in sync mode
4. common/lru: add test case for BasicLRU.Peek
5. core: log message if DeriveFields returns error
6. send big tx by announce only
7. make signature of ContractCode hash-independent
8. add cortex robot

市场运营

1.调研当前ZKML的生态和进展；
2.调研当前MPC钱包的技术进展；
3.NFT 艺术家合作对接；
4.持续搜集、翻译、原创深度相关内容，NFT及DeFi等领域，尝试开拓全新的社区、媒体内容运营方式；
5.持续对市场热门项目进行关注并分析项目逻辑及经济模型等。

| Website | GitHub | Twitter | Facebook | Reddit | Kakao | Mail | Discord

【周报】Cortex团队技术周报20230718 was originally published in Cortex Labs on Medium, where people are continuing the conversation by highlighting and responding to this story.

Cortex团队开发进展

MRT/CVM

1. ZKML 增加 TVM Runtime Inference 推理模拟测试量化后精度，并调试精度损失情况;

2. ZKML 调整数据集挂载接口和输出结果分析类;

3. ZKML 调整量化模型输出结果 restore 逻辑;

4. 修复 model2circom 中 input_data not passed mrt transformer 问题;

5. ZKML 调整数据集挂载，并在 calibrate pass 中增加数据集输入，提升量化精度;

6. Requantize bit位数调整，add/sub/mul/div 等二元算子调整为30/31，16/16;

ZkRollup

1. 通过Binaryen 和wasm-pack 实现zkx-crypto(签名，验证，hash等)相关的功能到wasm;

2. 添加batch tx，包括在mempool, statekeeper对其进行处理;

3. 实现batch 的fee计算，添加batch tx相关的api 接口，修复fee的计算问题;

4. Forced exit request 里面提交的tx 改成batch tx，确保同时成功或者失败;

5. 创建整个数据库的schema 和创建对应的index（sqlx migration）;

Cortex 全节点

1. [hash less func added](https://github.com/CortexFoundation/CortexTheseus/commit/087a1dced5c2e5c9dc4310ff678676c9b6aadd4c)

2. [narrow down the scope of the variable](https://github.com/CortexFoundation/CortexTheseus/commit/421bdef419f83219ce9bbd4b41f6c0b3b18a8b97)

3. [consensus/clique: use slices package for sorting (#27488)](https://github.com/CortexFoundation/CortexTheseus/commit/2cf451530acb68041bed1de2b509e3e4a0a7267a)

4. [metrics: use slices package for sorting (#27493)](https://github.com/CortexFoundation/CortexTheseus/commit/f0349835b1a4ed672cef7cd346e82687489f460d)

5. [cmd/devp2p: use slices package for sorting (#27487)](https://github.com/CortexFoundation/CortexTheseus/commit/185cc62c235a0736e860fa4125c8a6abd9dbf866)

6. [accounts/keystore: use slices package for sorting (#27485)](https://github.com/CortexFoundation/CortexTheseus/commit/c7acd960ac2440359d5e9923510798c5bc4d0af8)

市场运营

1.NFT艺术展馆正式上线发布；

2.NFT 艺术家合作对接；

3.持续搜集、翻译、原创深度相关内容，NFT及DeFi等领域，尝试开拓全新的社区、媒体内容运营方式；

4.持续对市场热门项目进行关注并分析项目逻辑及经济模型等。

| Website | GitHub | Twitter | Facebook | Reddit | Kakao | Mail | Discord

【周报】Cortex团队技术周报20230705 was originally published in Cortex Labs on Medium, where people are continuing the conversation by highlighting and responding to this story.