Kimi K3's 2.8T Open-Weights Drop, the Open-vs-Closed AI Security Fight, and Why a 10-Year-Old PC Just Ran a Modern LLM

The Trillion-Parameter Giant Actually Ships

Moonshot released the open weights for Kimi K3, a 2.8T-parameter mixture-of-experts model with roughly 104B active parameters, 896 experts (16 active per token), a 1M-token context window and native image understanding. It shipped alongside open-sourced attention kernels, an expert-communication library and agent-environment tooling, with a claimed ~2.5x scaling-efficiency gain over the previous generation. The weights are free, but the ~1.4TB checkpoint makes deployment brutal: eight top-tier data-centre GPUs still need multiple machines, one developer called it the first open model they can't run even on a 512GB Mac Studio, and a single-node rig was estimated near $500,000.

Nvidia's Open Secure AI Alliance

Jensen Huang formally launched the Open Secure AI Alliance, arguing that because attackers already wield strong AI, defenders need inspectable, open models rather than only black boxes. He pointed to a recent Hugging Face security incident where, he claims, an open-weight model helped contain the intrusion while a closed model blocked forensics. Members include Microsoft, IBM, Cloudflare, Cisco and Red Hat. OpenAI reportedly declined to join, prompting internal pushback.

Anthropic, Washington and the Politics of Open

After weeks of criticism, Anthropic clarified it has never sought to ban open-weight models, but supports chip controls on China, measures against industrial-scale distillation, and mandatory safety testing for highly capable systems. Separately, reporting suggests the US government may seek up to 30 days of pre-release access to frontier models for evaluation by agencies including the NSA, turning model releases into a governance checkpoint.

Nvidia Bets on Ilya Sutskever's SSI

Safe Superintelligence, Ilya Sutskever's secretive lab, announced a strategic partnership in which Nvidia will fund a roughly 10x compute expansion over the next 12 months on next-generation hardware, despite SSI having no shipping product.

AMD's Fully Open Instella

AMD released Instella, its first fully open mixture-of-experts model at 16B total / 2.8B active parameters, trained entirely on its own MI300X/MI325X accelerators. Unlike open-weights drops, it includes checkpoints across every training stage, data mixtures, configs and code, a complete reproducible artifact.

Microsoft's Streaming Vision Model

Microsoft shipped Mage-VL 4B, a compact codec-native streaming vision-language model built to understand live video as it happens rather than analysing static images after the fact.

The Reliability Tax on AI Agents

New research highlighted hidden fragility in AI agents: one evaluation ran agents through 227 sequential rounds of evolving requirements to catch silent regressions, while another found that across nearly 6,000 paired runs, adding agent skills improved new tasks but broke many previously solved ones, a genuine regression tax. A related result showed modules in multi-part RL systems drifting from their assigned roles.

Big Models on Old Hardware

A 10-year-old CPU-only server (Intel i7-6700, 32GB DDR4) ran a quantised ~35B MoE model at 5-10 tokens/second using aggressive low-bit quantisation. It fed a growing community push for deployable 30B-120B models over trillion-parameter giants, plus a finding that coding-harness overhead alone can swing the same model's runtime from about 2 minutes to 8.