A 2.8-trillion-parameter open model, an embedding checkpoint that just topped the retrieval leaderboard, an AI that out-hacks human red teamers 6-to-1, and xAI open-sourcing its coding agent. The frontier is moving in every direction at once — here's what shipped.

Moonshot dropped Kimi K3, an open-weight 2.8-trillion-parameter MoE with native multimodal reasoning and a 1M-token context window. The headline architecture is Kimi Delta Attention, which enables up to 6.3x faster decoding at million-token context — the mechanism that keeps a 2.8T/1M model from being unusably slow or ruinously expensive.

So what: This is the largest open-weight model shipped to date, and the real story is the attention trick — Delta Attention makes trillion-scale, million-token inference economically viable. Open labs aren't just matching frontier scale anymore; they're solving the cost curve that keeps big-context models off real workloads.

Voxtral Transcribe delivers the lowest word error rate of any transcription API. Speaker diarization, word-level timestamps, and context biasing across 13 languages, with recordings up to 3 hours per request. Need live audio? Voxtral Realtime streams sub-200ms, open-weights under Apache 2.0 and deployable on edge.

Try Voxtral now

_sponsored

NVIDIA open-sourced Nemotron-3-Embed, a multilingual embedding collection (1B and 8B). The Nemotron-3-Embed-8B checkpoint ranks #1 overall on RTEB (the Retrieval Embedding Benchmark) with a score of 78.5, built on Ministral-3-8B and mapping text into 4096-dim dense vectors. It tops the leaderboard while lowering token spend across tasks.

So what: Retrieval is the memory of every agent — better embeddings compound into better RAG, better tools, better long-horizon behavior. Winning RTEB and cutting token cost means this is the new default retrieval backbone for enterprise agent stacks, and it's open.

OpenAI detailed GPT-Red, an internal self-play red-teaming model trained to find prompt-injection weaknesses at scale. On novel scenarios it hit an 84% attack success rate vs. 13% for human red-teamers. Using it to harden GPT-5.6 Sol produced ~6x fewer direct injection failures than GPT-5.5.

So what: Security is becoming a self-play loop — an attacker model trained to break the defender, then folded back into training. As agents get real system access, prompt injection is the attack surface that matters, and automated red-teaming at 84% success is how you find the holes before someone else does.

SpaceXAI open-sourced Grok Build — the terminal coding agent behind its grok CLI — under Apache 2.0. Crucially, it's the harness, not the model: the Rust source for the full-screen TUI, the agent runtime, and the tool layer (file edits, shell execution). It runs interactively, headlessly in scripts, or via the Agent Client Protocol (ACP), and syncs periodically from the internal SpaceXAI monorepo.

So what: Open-sourcing the harness under a permissive license decouples the agent scaffold from Grok's weights — anyone can now run xAI's exact coding-agent runtime on any model.