Gemma 4: Google's Most Capable Open Models Under Apache 2.0
Google DeepMind released Gemma 4 on April 2, 2026, as the most capable open model family the company has shipped to date. The headline change is the license: Gemma 4 is the first Gemma release under the OSI-approved Apache 2.0 license, ending the custom license restrictions that previously complicated commercial deployment and redistribution. The Gemmaverse community has downloaded earlier Gemma models over 400 million times and produced more than 100,000 derivative models, so the shift to Apache 2.0 removes a legal friction point that has blocked enterprise adoption.
The family ships in four sizes to cover the full deployment spectrum. The edge-optimised E2B (~2.3B effective parameters) and E4B (~4.5B effective parameters) models target mobile and on-device inference. The 26B Mixture-of-Experts model activates only ~4B parameters per forward pass, delivering mid-tier capability with low per-token compute cost. The 31B dense flagship currently ranks third among open models on the Arena AI text leaderboard. All models support up to 256K token context windows, 140+ languages, and native multimodal input combining text and images. Smaller models add audio; video input is supported via frame extraction.
Gemma 4 is purpose-built for agentic workflows. Every size ships with multi-step reasoning, structured function calling, and thinking modes that allow the model to emit a scratchpad before producing its final answer — making it suitable for tool-using agents without fine-tuning. The models are available immediately on Hugging Face, Google AI Studio, Kaggle, and Ollama, and are running in preview on Vertex AI.
Read more — Google Open Source Blog
GitHub Copilot April 2026: Critic Agent, Public SDK, BYOK, and Signed Commits
GitHub shipped a dense set of Copilot updates across the first week of April 2026, advancing the platform on four distinct fronts.
The most significant new capability is the Critic agent, introduced on April 4. When Copilot completes a plan or a complex multi-step implementation, the Critic agent independently reviews the output using a complementary model — separate from the one that produced the plan — catching logical errors, missing edge cases, and incomplete steps before the developer sees them. This mirrors review-then-revise patterns used in multi-agent research and addresses the failure mode where a single model confidently produces plausible but incorrect plans.
The Copilot SDK entered public preview on April 2, providing building blocks for embedding Copilot's agentic capabilities directly into custom applications, workflows, and internal platform services. Developers can use the SDK to invoke Copilot agents programmatically, surface them inside internal tools, or chain them with proprietary context not available to the cloud agent.
On April 7, Copilot CLI gained support for BYOK and local models, letting teams route CLI completions through their own model provider or a locally running model instead of GitHub's hosted inference. Organization admins received firewall and runner controls on April 3 — firewall rules limit the cloud agent's internet access and protect against prompt injection; custom runner settings let organisations route agent development environments through private GitHub Actions infrastructure. Every commit made by the Copilot cloud agent is now cryptographically signed and marked as Verified on GitHub.
Also on April 7, Dependabot alerts became assignable to AI agents for automated remediation — moving beyond version-bump PRs to allow agents to reason about the vulnerability, test a fix, and open a pull request with context.
Read more — GitHub Changelog
OpenAI Safety Fellowship: Funded External Alignment Research Program
OpenAI announced the OpenAI Safety Fellowship on April 6, 2026, opening a structured program for external researchers to pursue safety and alignment work on advanced AI systems. The fellowship runs from September 14, 2026 through February 5, 2027, and provides participants with a $3,850 weekly stipend alongside approximately $15,000 per month in compute resources — enough to run substantial empirical experiments without access to institutional GPU clusters.
The program targets seven priority areas: safety evaluation, ethics, robustness, scalable mitigations, privacy-preserving safety methods, agentic oversight, and high-severity misuse domains. Fellows are expected to produce a concrete research artefact by program end — a paper, benchmark, or dataset — rather than a general survey or position paper. Applications close May 3, 2026, with successful candidates notified by July 25.
The fellowship is notable for funding independent researchers rather than OpenAI employees, which addresses a recurring criticism that alignment research published by frontier labs is structurally incentivised toward findings that do not constrain deployment timelines. The program comes at a moment when the broader industry is accelerating agentic deployments, and questions about oversight mechanisms for long-running agents are moving from theoretical to operational.
Read more — OpenAI
Links & Sources
You May Also Enjoy
