Moonshot launches Kimi K3 (open 2.8T), NVIDIA announces Open Secure AI Alliance, Perplexity beefs up its enterprise offering

July 27, 2026 is a packed day, with three distinct major announcements. Moonshot AI releases the weights and technical report for Kimi K3, an open 2.8-trillion-parameter model hosted from day 0 by Together AI. NVIDIA brings together around thirty partners that same day — from Adobe to Microsoft, via Hugging Face and SpaceXAI — under the Open Secure AI Alliance, a coalition for open tools for AI security, following the already covered Hugging Face security incident. Perplexity, for its part, publishes its first changelog in two weeks, a burst of nine highly enterprise-oriented features. Around these three topics revolve seven notable updates — Cognizant and Anthropic, Cohere North Automations, Zed’s adoption of Opus 5 — and five smaller announcements.

Kimi K3: Moonshot AI publishes an open 2.8-trillion-parameter model, hosted from day 0 by Together AI

July 27, 5:14 PM — Moonshot AI releases the weights and technical report for Kimi K3, described by the company as the first open model in the 3-trillion-parameter class — a Moonshot formulation, to be taken as such. It is a MoE (Mixture-of-Experts) model with 2.8 trillion total parameters, of which only 104 billion are activated per token, spread across 93 layers. Two in-house architectural innovations account for most of the claimed gains: Kimi Delta Attention (KDA), combined with an Attention Residuals mechanism, and a highly sparse MoE (“Stable LatentMoE”) that activates only 16 experts out of 896 on each pass. Moonshot claims roughly a 2.5× gain in scaling efficiency compared with Kimi K2 — “not just more parameters.” The model also natively integrates vision (MoonViT-V2 encoder) and a one-million-token context window.

Kimi K3 is available immediately on Kimi.com, the mobile app (iOS/Android/HarmonyOS), the Kimi Work desktop app, the Kimi Code command-line coding agent (selectable via /model) and the Kimi API (model kimi-k3, OpenAI/Anthropic compatible), priced at $0.30/MTok input with cache,$ 3.00/MTok without cache, and $15.00/MTok output. Together AI is a hosting partner from day 0, and Hugging Face reposts the announcement without additional comment.

Benchmark	Kimi K3	Claude Fable 5	Claude Opus 4.8	GPT-5.5
SWE-Marathon	42.0	35.0	40.0	—
MCPMark-Verified	94.5	87.4	—	—
GPQA Diamond	93.5	92.6	91.0	93.5

Moonshot explicitly acknowledges several limitations: strong sensitivity to the preserved reasoning mode (the full history must be passed back exactly as is in subsequent turns, or instability may occur), a tendency toward “excessive proactivity” on long and ambiguous tasks, and a user-experience gap described as “notable” by the lab itself versus Claude Fable 5 and GPT-5.6 Sol. A prefix cache implementation compatible with the new attention mechanism was also contributed to the vLLM community alongside the model.

“Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window. New model architecture: 2.5x the intelligence per unit of compute, not just more params.” — @Kimi_Moonshot on X

🔗 Model weights — Hugging Face

🔗 Technical report — GitHub

NVIDIA brings together the Open Secure AI Alliance, around thirty partners for AI security

July 27 — NVIDIA announces the creation of the Open Secure AI Alliance, a coalition of around thirty founding partners — including Adobe, Cadence, Capital One, Cisco, Cloudflare, Cognition, CrowdStrike, Databricks, Dell, Hugging Face, IBM, LangChain, the Linux Foundation, Microsoft, NAVER, Palantir, Red Hat, Salesforce, SAP, ServiceNow, Siemens, Snowflake, SpaceXAI and Thinking Machines Lab — dedicated to developing and sharing open tools for the security and reliability of AI systems.

The announcement explicitly builds on the security incident that occurred at Hugging Face in mid-July (already covered): NVIDIA’s blog notes that Hugging Face ran the open model GLM 5.2 on its own infrastructure to analyze more than 17,000 actions and contain the intrusion — proof, according to NVIDIA, that a state-of-the-art open model is a defensive asset, not a security liability.

Alliance partner	Concrete contribution
NVIDIA	NOOA (NVIDIA Labs Object-Oriented Agent), a research framework for agent harnesses, open source on GitHub
Hugging Face	Donation of the Safetensors format to the PyTorch Foundation
Microsoft	MDASH, multi-model agentic scan harness to prove exploitable flaws
HPE	SPIFFE/SPIRE, zero-trust identity for AI agents and services
SpaceXAI	Release of the Grok Build terminal coding agent, Grok lineage weights to come

Unusually, Jensen Huang comments on the announcement himself — only his second post ever published on X, after the July 24 letter already covered. The same day, Cognition (Devin) joins the alliance by contributing its own research on the reliability of models derived from open source, already published on July 8:

“Closing open source models hurts innovation. The path forward is better tools to evaluate, secure, and deploy them responsibly.” — @cognition on X

Sakana AI joins the open-model movement the same day, following the Jensen Huang letter of July 24 — a third signal, distinct from the alliance itself, that closes out the week: security incident (July 21), open letter (July 24-26), concrete industry alliance (July 27).

🔗 Official announcement — NVIDIA

🔗 Jensen Huang on X

🔗 Sakana AI joins the movement

Perplexity publishes its first changelog in two weeks: nine features, enterprise focus

July 27 — Perplexity publishes a dense changelog entry, the first since July 13 (“Introducing Brain for Perplexity Computer”), after two weeks with no new post on this channel. The entry bundles nine distinct features touching Perplexity Search, Computer, the Agent API and Comet Assistant. Eight are new; the ninth — the integration of Claude Opus 5 into Search and Computer — had already been announced on X on July 25 and is only officially confirmed here in the changelog.

New feature	Technical detail
Role-based access control (RBAC) + SCIM	Custom roles, group synchronization, per-group credit limits — reserved for annual Enterprise contracts
API credentials vault	Computer authenticates to any third-party API without exposing the secret to the agent or the sandbox
Brain expanded to Max subscribers	Self-improving memory extended to all Max subscribers worldwide, wiki writing in 15 languages
Skills for the Agent API	Composable capabilities loaded on demand for developer agents
Computer in Comet Assistant	Access to Computer directly from the Comet panel, without leaving the page being viewed
Check Sources	Verifies that an Ask or Computer answer is actually supported by its cited sources
Source Context Panel	Side panel to inspect, compare, and dig deeper into a response’s sources
Prompt-based session management	Retrieve, summarize, fork, or parallelize Computer sessions from a simple conversation

The first two features in the table specifically target organizations under annual Enterprise contract: role-based access control with SCIM synchronization addresses a classic IT governance need (centralized provisioning, team credit limits), while the API credentials vault extends Computer to third-party services without a native connector, without ever exposing the raw secret to the agent. The expansion of Brain to all Max subscribers, meanwhile, affects a much broader audience than the Enterprise segment alone, where the feature had started a month earlier.

None of these nine features is flagged as being limited to a preview: all are deployed or rolling out generally at the time of publication. By its scope — enterprise governance, developers and the general public in the same entry — this changelog is Perplexity’s densest publication since the launch of Personal Computer in mid-July.

🔗 Official changelog — Perplexity

Cognizant and Anthropic expand their enterprise partnership

July 27 — Cognizant becomes a Global Premier Partner within the Claude Partner Network. Claude is now integrated into three internal platforms at the company: Flowsource (full-stack engineering platform, Spec-Driven Development module where Claude Code is guided by the project’s specifications and blueprints), Neuro AI Engineering and Neuro IT Ops. More than 30,000 Cognizant associates have already been trained on Claude, as part of the group’s new “Frontier Certified” workforce model.

Customer case	Measured result
Customer experience portal (global manufacturer)	Delivered in 6 months
Contract intelligence (biopharma)	Up to -40% review time, over 88% extraction accuracy
Risk navigation (insurance)	Around 8h/week/person saved

Ravi Kumar S (Cognizant CEO) points to a growing gap between AI capabilities and companies’ ability to absorb them, positioning Cognizant as the bridge providing the business context, engineering scale and trust frameworks needed for these deployments.

🔗 Official announcement — Anthropic

Coding tools: Zed adopts Claude Opus 5, GitHub Copilot optimizes its costs

Zed 1.12.1 and 1.13.1 add Claude Opus 5

July 27 — Zed simultaneously releases versions 1.12.1 (stable) and 1.13.1 (preview), adding support for Claude Opus 5 for Anthropic and Amazon Bedrock BYOK (bring your own key) providers. Zed becomes the fifth coding tool to integrate Opus 5, after Cursor, Devin, v0 and Warp — but, unlike the previous four, only via changelog, with no announcement on X. The preview channel also fixes three distinct bugs: project settings not re-enabling language servers, duplicated responses with GPT-5.x models on Amazon Bedrock’s Mantle endpoint, and a project search surfacing nested repositories that were supposed to be excluded.

🔗 Stable changelog — Zed

GitHub Copilot — Auto routing (HyDRA) makes AI credits last 3.3× longer

July 27 — GitHub explains that prompt caching, tool search and a model-routing system called HyDRA reduce repeated context on each Copilot session. The numerical result: HyDRA matches the 70.8% task resolution rate achieved by OpenRouter Auto, for a 3.3× gain in credit savings. No dedicated blog post has been published for this announcement, which so far remains a simple tweet — in line with the Copilot harness efficiency evaluation published on June 25.

🔗 Announcement — @github on X

OpenAI: voice expands to enterprise and new economic research

GPT-Live available globally for Edu, Business and Enterprise

July 27 — OpenAI confirms that GPT-Live, the new generation of voice models introduced on July 8 for natural human-AI interaction, is now available worldwide for Edu, Business and Enterprise plans. The ChatGPT Voice desktop app had already mentioned these plans on July 23; the novelty here is the explicit confirmation that the voice model itself — not just the desktop integration — is covered, globally. The announcement is posted directly on X, with no dedicated article: it is the element most directly tied to the Enterprise expansion observed in this monitoring window.

🔗 Announcement — @OpenAI on X

Work at the Frontier: how AI expands tasks at work

July 27 — OpenAI Economic Research publishes the first report in a new recurring series, based on the analysis of more than 800,000 ChatGPT messages from U.S. users. The report introduces the notion of task crossover: work usually associated with one occupation that appears in the AI usage of people in another occupation. 16.8% of work-related messages, and 43.5% of occupation-specific messages, concern tasks from a different occupation than the user’s.

Occupation concerned	Share of out-of-occupation tasks
Customer experience	77%
Design	75%
Human resources	69%
Legal	56%
Marketing	53%

The phenomenon is more frequent in smaller organizations: the share of out-of-occupation tasks falls from 18.9% in workspaces with 2 to 5 seats to 16.3% beyond 100 seats, consistent with the idea that AI serves as a general-purpose tool where specialized resources are lacking.

🔗 Full report — OpenAI

Meta AI deploys SAM and DINOv3 in an assistive robotics project

July 27 — Meta AI publishes a case study on the RAMMP project (Robotic Assistive Mobility and Manipulation Platform), led by the University of Pittsburgh (HERL) and ATDev, funded up to $41.5 million by ARPA-H (the U.S. federal agency for advanced biomedical research). The goal: safer robotic platforms for wheelchair users — around 5.5 million people in the United States, with more than 100,000 wheelchair-related injuries treated every year in emergency rooms.

The open-source models DINO/DINOv3 (self-supervised vision transformer) and SAM/SAM 3.1 (segmentation) run in edge computing directly on the robots’ onboard hardware — a constraint imposed by battery autonomy and the lack of reliable network connectivity. The perception system relies on RF-DETR, a lightweight detector fine-tuned with DINOv2 embeddings, whose training data are auto-annotated via SAM. The associated consortium includes Kinova Robotics, LUCI Mobility, ATDev, and teams from Carnegie Mellon, Cornell, Northeastern, and Purdue.

🔗 Full case study — Meta AI

Cohere launches North Automations, agentic workflow orchestration for the enterprise

July 27 — Cohere announces North Automations, a new capability of its enterprise agentic platform North, making it possible to design automated workflows in natural language: scheduled runs, loops and branches, model choice by step, a “Plan” mode to review before building, human approval points, and token consumption tracking. Cohere puts the market opportunity for agent orchestration at $550 billion, citing a Gartner report on the agentic AI hype cycle.

Automations builds on North’s existing infrastructure (on-prem or cloud deployment, proprietary connectors, MCP or SDK, guardrails and granular permissions) and is available today to all North customers. Internally, the Marketing Operations, Customer Success, and Sales teams have already built their own automations — respectively a natural-language translation of questions into BigQuery SQL queries, a customer engagement tracker aggregating HR, Slack, and Salesforce, and a prospect call prep flow combining Salesforce, Slack, Gmail, and Notion.

🔗 Announcement — @cohere on X

🔗 Blog post — Cohere

Briefs

GitHub Blog — “The harness is all you need (mostly)” — Burke Holland publishes an 8-step Copilot workflow (choose a tool, YOLO mode, visual prototype, plan with /grill-me, implementation by Autopilot, human review, cross-review by a model from another family, delivery), valid across the entire Copilot family. 🔗 Source
NVIDIA — Vera CPU speeds up chip design — deployed internally in EDA flows with Cadence (Jasper) and Synopsys (VCS), up to 1.5× more performance on certain verification workloads, with the same number of cores. 🔗 Source
NVIDIA — Nemotron 3 Ultra leads in agentic RTL coding — across nine categories of real chip design, this already well-known open model achieves the best results among the open models tested. 🔗 Source
HeyGen / HyperFrames — days 21 and 22 of 30 — cloud rendering via the HeyGen API (around $0.10/minute at 1080p) then a catalog of reusable components to build scenes without starting from scratch. 🔗 Day 21 · Day 22
ElevenLabs — summit in Bengaluru on October 6 — previews of new models and demonstrations from Indian teams working on voice AI announced for this event. 🔗 Source

What this means

Model openness is moving from words to action. After Jensen Huang’s July 24 letter and the growing rally behind it (Cohere, Ai2, Together AI, then Sakana AI this weekend), July 27 marks a qualitative leap: a structured industrial alliance with concrete tools contributed by its members (NVIDIA’s NOOA, Hugging Face’s Safetensors, HPE’s SPIFFE/SPIRE), and a flagship model, Kimi K3, claiming the 3 trillion-parameter class in open weights. Two different forms of the same dynamic, published on the same day.

Security as an argument for openness, rather than against it. The Open Secure AI Alliance turns a common argument around: it is not model openness that weakens security, but their closure that deprives defenders of inspectable tools. The mid-July Hugging Face incident, where an open model (GLM 5.2) enabled forensic analysis of more than 17,000 intrusion actions, serves as a demonstration of this thesis — a textbook case that NVIDIA explicitly cites to justify creating the alliance.

The open frontier is getting closer to the closed frontier. On several precise agentic benchmarks — SWE-Marathon, MCPMark-Verified, GPQA Diamond — Kimi K3 surpasses or matches Claude Fable 5, Claude Opus 4.8, and GPT-5.5, while itself acknowledging that it still trails in overall user experience. The gap between open and proprietary models is narrowing benchmark by benchmark, even if open labs themselves admit they do not yet match the overall experience of the best closed models.

Enterprise AI is being industrialized through small governance building blocks. Cognizant training 30,000 associates on Claude, Perplexity adding RBAC and SCIM to Computer, Cohere launching workflow orchestration with human approval points, OpenAI extending GPT-Live to Edu/Business/Enterprise: four different players are moving on the same day toward building blocks for access control, traceability, and automation — the foundation needed for agentic AI to move from prototype to large-scale enterprise deployment.