Gemma 4 12B Brings Multimodal Agents to Your Laptop

Google's new open-weights model runs full audio+vision+text inference on 16GB of RAM, no cloud required.

Why it matters: Encoder-free architectures have been a research bet — Gemma 4 12B ships one in production, meaning faster inference, lower memory, and a single backbone for all modalities. For teams building local or privacy-sensitive agentic workflows, the hardware bar just dropped significantly.

• Ships under Apache 2.0 with Multi-Token Prediction drafters to cut latency; runs via Ollama, LM Studio, llama.cpp, and vLLM out of the box.
• Benchmark performance tracks close to the 26B MoE model at less than half the memory footprint; first mid-sized Gemma model with native audio input.

Go deeper: https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12b/

Mnemo: Persistent Memory for Any LLM in Under 50ms

A new open-source sidecar gives every LLM a knowledge graph that survives session boundaries.

Why it matters: Stateless LLM calls are the core pain point in long-running agents and assistants. Mnemo solves this with a local Rust service that extracts entities, builds a SQLite+petgraph knowledge graph, and injects context back into prompts — no cloud, no vendor lock-in.

• Works with Ollama, OpenAI, Anthropic, or any OpenAI-compatible backend; ships as a single static binary with a REST API, CLI, and Python SDK.
• Full retrieval pipeline benchmarks at ~4.2ms average on M2; entity insert runs at ~8,300 ops/s in SQLite WAL mode.

Go deeper: https://github.com/zaydmulani09/mnemo

Let's Encrypt Plans Post-Quantum Certs by 2027 via Merkle Tree Approach

Let's Encrypt is betting on Merkle Tree Certificates (MTCs) to add post-quantum authentication without bloating TLS handshakes.

Why it matters: Dropping in ML-DSA signatures naively would push a TLS handshake past 10KB — Cloudflare research shows that breaks real-world connections at scale. MTCs batch-sign certificates and keep the per-connection authentication path smaller than today's handshakes, even with post-quantum algorithms. Chrome has already named MTCs its preferred path.

• Staging environment targeting late 2026; production in 2027. Current certificates continue unchanged.
• If you operate servers today, the high-leverage move is enabling hybrid post-quantum key exchange (X25519MLKEM768) — major browsers and OSes already support it on the client side.

Go deeper: https://letsencrypt.org/2026/06/03/pq-certs.html