2026-07-27

1. Kimi K3 open weights released today — world's largest open model, data sovereignty angle

Score: ~1.7 · Open-weight × Frontier models · +weights bonus

Moonshot AI released the full weights for Kimi K3 on July 27 as promised at launch (July 16). License: Modified MIT (commercial use permitted). Hosted on Hugging Face (moonshot-ai/kimi-k3). Precision: MXFP4, requiring ~1.4 TB of fast memory for full inference.

Why it matters: This is the largest model ever publicly released at open weights — 2.8T parameter MoE (896 experts, ~16 active per token). The Modified MIT license enables commercial fine-tuning and deployment by any party. The data sovereignty angle is real: self-hosted inference doesn't touch Moonshot servers, making it the first frontier-class Chinese model that US/EU enterprises can run without data routing concerns about Chinese API infrastructure. Yesterday's Watch item is resolved: no US sanctions action materialized before the weights dropped. → Kimi K3, Moonshot AI

Source: Hugging Face model card (source)

2. DiffusionGemma — first open-weight text diffusion model from a major lab (⚡ day+46

Score: ~1.8 · Frontier models × Google DeepMind × new page · ⚡ day+46 from June 11 release

Google DeepMind released DiffusionGemma on June 11, 2026 — a 26B MoE model built on the Gemma 4 26B-A4B backbone. This is a, not left-to-right autoregressive decoding. Speed: 1,000+ tokens/sec on a single H100 GPU, approximately 4× faster than autoregressive Gemma 4 26B-A4B. Memory: ~18 GB VRAM (MXFP4). Apache 2.0.

Accuracy tradeoff: AIME 2026 drops from 88.3% to 69.1%; MMMU Pro drops from 73.8% to 54.3%. Multi-step reasoning takes the hardest hit; pure text tasks fare better.

Why it matters: DiffusionGemma is the first open-weight production text diffusion model from a major AI lab. It establishes a new engineering knob: operators can now trade ~19pp of reasoning accuracy for 4× throughput on high-volume text generation tasks. Autoregressive was the only viable decode path at scale; it no longer is. Six weeks in, the ecosystem has had time to start building on it — but this system hadn't tracked it until today. → DiffusionGemma, Google DeepMind

Source: DeepMind model page (source)

3. Grok enters Google Workspace (Sheets, Slides, Docs) — cross-platform productivity strategy complete

Score: ~1.1 · Agents/tool use × xAI · product feature

SpaceXAI launched a free Grok add-on for Google Workspace on July 24 (via Google Workspace Marketplace). Grok appears as a sidebar panel in Sheets, Slides, and Docs. Key capability in Sheets: reads selected ranges, cites specific cells behind its answers, writes and fills formulas, rebuilds financial scenarios. One week prior (July 17–18), SpaceXAI shipped the Microsoft Office (Excel/Word/PowerPoint) add-in.

Why it matters: Grok now spans both major office suites simultaneously. ChatGPT has Copilot (Microsoft) but not Google Workspace; Gemini has Google Workspace but not Office. xAI's cross-platform strategy positions Grok as the only frontier AI sidebar available inside both ecosystems at once — and at free pricing, which removes the adoption friction. Combined with Grok Automations (July 16) and Grok Connectors, the agent surface area keeps growing: conversational interface → scheduled tasks → email triggers → office-suite sidebar in less than two months. → xAI

Source: SpaceXAI blog (source)