How to Verify an AI Model's Weights Haven't Been Tampered With

A step-by-step guide to independently verifying that a downloaded AI model checkpoint matches what its publisher actually released.

OpenAI Comes Back to Open Weights, Six Years Later

GPT-OSS-20b marked OpenAI's return to open-weight releases after nearly six years of closed, API-only models.

Can You Prove When an AI Model Was Released?

Why release dates and knowledge cutoffs for AI models are harder to verify than they look, and what it actually takes to prove one independently.

What Is AI Model Provenance, and Why Does It Matter?

Provenance means being able to verify an artifact's origin and whether it has been altered since. Here's what that means for AI model weights specifically.

The Weekend DeepSeek-R1 Shook the AI Market

DeepSeek-R1 was among the first open models to rival closed frontier reasoning models -- and its release triggered an immediate, widely reported market reaction.

DeepSeek-V3 Quietly Rewrote the Rules on Training Cost

DeepSeek-V3 claimed frontier-competitive performance at a training cost far below prior industry assumptions, setting up DeepSeek-R1's release weeks later.

Inside Alibaba's Relentless Qwen Release Cadence

Alibaba's rapid-cadence Qwen family made Qwen2.5-7B one of the most widely fine-tuned and deployed open models of its generation.

Open Weights vs. Open Source AI: What's the Difference?

"Open" AI models vary widely in what's actually public. Here's the difference between open weights, open source AI, and permissive licensing.

What Is EAS (Ethereum Attestation Service)?

An introduction to the Ethereum Attestation Service (EAS), the open protocol Weight Registry uses to record verifiable, structured claims on-chain.

Mixtral-8x22B and the Rise of Open Mixture-of-Experts Models

Mistral AI scaled its sparse Mixture-of-Experts architecture to 141B total parameters and released it fully open under Apache 2.0.

What Is a Model Weight Hash, and Why Does It Matter?

A plain-language explanation of what a cryptographic hash of an AI model's weights actually is, and why it matters for verifying a checkpoint hasn't been altered.

The 7B Model That Out-Punched Its Weight Class

Mistral AI's first model matched a competitor nearly twice its size, under a fully permissive license -- and became a default base for open fine-tuning.

How Vicuna Turned ChatGPT Logs Into an Open-Source Chatbot Boom

LMSYS's Vicuna helped popularize fine-tuning open base models on real conversational data -- and the Chatbot Arena evaluation method that came with it.

Inside BLOOM, the Year-Long Experiment in Open, Multilingual AI

BLOOM was the product of a year-long international research collaboration -- and one of the largest genuinely open multilingual models of its time.

GPT-J-6B: The Open Answer to a Closed GPT-3

EleutherAI's GPT-J-6B proved GPT-3-class capability didn't require a closed API, years before open weights became a mainstream strategy.

Before ChatGPT, There Was GPT-1

OpenAI's original GPT, released in 2018, set the template every subsequent GPT release -- and much of the field -- would follow.