Open Source AI Pulse

New evaluations expose how benchmarking methods now shape results as much as models themselves.

• Coding performance gaps remain wide: Claude Opus 5...

Anthropic faces simultaneous open-source pushback on two fronts, exposing safety-open source tensions.

• Regulatory capture claims: David Sacks...

Chinese open-weight models are shifting from niche experiments to practical production choices for builders facing high inference costs.

-...

Open-source projects face a surge in low-quality AI code submissions that shift verification work onto already stretched maintainers.

• Home...

Nvidia is investing $1 billion and paying $6 billion to license Poolside's tech, hiring over 100 engineers to advance its Nemotron open-weight models....

A new study found all 21 tested open-weight LLMs could have their safety guardrails stripped despite built-in protections, raising risks of scaled...

Framework choice now drives production success more than ever, with most teams failing to scale due to mismatches.

• LangChain/LangGraph leads in...

Industry giants from SoftBank, Hitachi, NTT, Google, Microsoft, and Anthropic will gather in Tokyo at AGNTCon and MCPCon Japan to set standards for...

This week's trends highlight shifting AI dynamics:

• NY overtakes SF in tech talent with 394k workers vs 376k, driven by finance AI hiring
• Qwen 3.8...

• Partnership taps Lancium's 15+ GW pipeline and 4 GW leased capacity for full NVIDIA AI factory deployments
• NVIDIA strategic investment backs...

• lemans releases as open-source Ruby framework with guardrails like snapshot restoration and credential filtering for safe agent eval
• Nvidia AVO...

An unnamed provider dropped Ox Alpha on OpenRouter, describing it as a reasoning model for coding and agentic work. Early benchmarks show it rivaling...

Hugging Face's potential acquisition threatens its role as the trusted neutral hub for open weights, as buyer ownership by model providers could erode...

TTOs treat open-source vs. traditional licensing as a fit question, not a binary. They weigh four practical factors before deciding:

• Value location:...

Two major efforts are tackling what "open" really means for AI models beyond code.

• Linux Foundation's OpenMDW license, backed by Meta, Amazon, IBM,...

IBM and Together AI's $240M multi-year deal deploys the first large-scale NVIDIA HGX B300 inference cluster on IBM Cloud in Q1 2027, giving Together AI the infrastructure to deliver faster, lower-cost open-source model inference for enterprises.

Gemma models hit 1 billion downloads just two years after launch, powering production deployments far beyond benchmarks.

• Space: NASA, Satlyt and...

The Berkeley Humanoid Lite lets anyone build a functional humanoid using home 3D printers and under $5,000 in standard parts, with fully open-source...

Anthropic is expanding access to its Claude Mythos 5 cybersecurity model for more defenders while launching a $35M open-source security fund. The moves aim to solidify its competitive position in AI cybersecurity and support valuation growth.

Training AI on copyrighted books can qualify as fair use when transformative and non-competitive, but pirating the source material remains illegal, as...