OpenAI has put conversations and Projects back in its redesigned ChatGPT desktop app and enabled cloud Work threads to move across devices, correcting the launch's biggest usability failures without merging local Work or Codex histories.
Moonshot AI has made Kimi K3 available through its apps and API, pairing a 2.8-trillion-parameter architecture with early frontier-level results, but the model's open-weight claim cannot be tested until its weights and technical report arrive.
Moonshot AI has started selling access to a 2.8-trillion-parameter model before releasing the weights needed to inspect, modify or operate it independently. Kimi K3 may prove to be a consequential open model, but at launch it is a hosted frontier service carrying an open-weight promise.
Moonshot says in its announcement that K3 combines native vision, a 1-million-token context window and a mixture-of-experts architecture built for long coding, reasoning and knowledge-work tasks. It is available through Kimi's chatbot, Work and Code products and the Kimi API.
The company says it will release the full weights by July 27, together with a technical report covering the architecture, training and evaluations. Until then, customers can use K3 on Moonshot's infrastructure but cannot independently inspect, modify or run the promised release. Early descriptions of K3 as already open collapse that distinction.
The headline parameter count also needs qualification. Moonshot calls K3 the first open model to reach 2.8 trillion total parameters, compared with 1.6 trillion reported for DeepSeek V4 Pro and 744 billion for Zhipu AI's GLM 5 series. K3 is sparse: Moonshot says it effectively activates 16 of 896 experts. The 2.8-trillion figure describes total capacity, not the number of parameters used for every token.
Moonshot attributes an approximately 2.5-fold gain in scaling efficiency over Kimi K2 to Kimi Delta Attention, Attention Residuals, Stable LatentMoE and changes to its data and training recipes. Those claims remain company measurements pending the technical report and outside replication.
The infrastructure guidance is a second limit on the practical meaning of openness. Moonshot recommends supernodes with 64 or more accelerators and says a compatible vLLM implementation for Kimi Delta Attention will arrive with the weights. A downloadable checkpoint could reduce dependence on Moonshot's API, but serving K3 will still require substantial hardware and engineering.
K3's strongest independent result is specific. In blind human testing, it ranked first on Arena's frontend-coding leaderboard and beat Claude Fable 5 in five listed domains while trailing it in gaming, according to the launch coverage.
On the broader Artificial Analysis Intelligence Index, K3 scored 57. That placed it around Claude Opus 4.8 and GPT-5.5 but behind Claude Fable 5 and GPT-5.6 Sol. Moonshot also says K3's overall performance still trails those two leading proprietary systems.
Moonshot's own benchmark table contains real wins, but not a uniform lead. K3 beat Fable 5 and Sol on the reported Program Bench and SWE Marathon results, while trailing one or both on DeepSWE, FrontierSWE, GDPval-AA v2 and several reasoning tests. The comparisons also mix Kimi Code, Claude Code and Codex harnesses. K3 ran at maximum reasoning effort, and Moonshot says some Fable 5 results may include fallback behavior.
That matters for both performance and economics. K3 launches with maximum thinking effort as the default; lower-effort modes are promised in later updates. The available results therefore do not establish how cheaper or faster settings will perform.
Moonshot lists further product limitations: K3 can become unstable when an agent harness fails to preserve its thinking history, may take unexpected action when user intent is ambiguous, and has a noticeable user-experience gap versus Fable 5 and Sol. An early assessment adds the basic launch-week caveat: K3 had been available for only hours, so initial benchmarks and viral demonstrations may overstate its reliability in sustained work.
Moonshot charges $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens and $15 per million output tokens. It says its own API achieves a cache-hit rate above 90% on coding workloads, a company figure whose relevance will depend on how often a customer's prompts can be reused.
The output price is much higher than those of lower-cost Chinese open-weight peers. Artificial Analysis nevertheless estimated K3's cost per task at $0.94, just below GPT-5.6 Sol at $1.04 and about half Claude Opus 4.8 at $1.80. Those measures answer different questions: token pricing is a rate card, while cost per task depends on how much work the model consumes. Neither includes the hardware and engineering cost of self-hosting.
Open release is not a new tactic for Moonshot. When the company released Kimi K2 in July 2025, a contemporaneous account described open models as a way to demonstrate technical capability, build developer communities and extend global influence. It also recorded the pressure behind that strategy: after DeepSeek's low-cost models rose, the Kimi app fell from third in monthly active users in August 2024 to seventh by June 2025, according to the tracker cited in that account.
K3 therefore enters a crowded field that includes DeepSeek and Zhipu AI as well as models from Alibaba, Tencent and Baidu. Its release had an immediate market effect—Zhipu's Hong Kong-listed shares fell as much as 25% by Friday midday—but that price move reflects changing investor expectations, not independent validation of K3's performance.
Moonshot has since attracted substantial new financing. The company, founded in 2023 by former Meta AI and Google Brain researcher Yang Zhilin, raised about $2 billion at a $20 billion valuation in May. A report based on information from its financial adviser said Moonshot had raised $3.9 billion over the preceding six months and that annual recurring revenue exceeded $200 million in April, driven by subscriptions and API use.
Those figures show the commercial logic behind a two-track launch. Moonshot can sell hosted access immediately, while a later weight release can broaden adoption among developers and infrastructure providers. They do not show how much of the new capital K3 consumed or whether its economics will match smaller alternatives.
The next consequential evidence is not another launch-week leaderboard. It is whether Moonshot publishes the full weights and technical report by July 27, with license terms, inference code and documentation that allow meaningful inspection, modification and deployment.
After that, independent operators will need to answer three questions: whether the sparse 2.8-trillion-parameter system can be served reliably outside Moonshot; whether production users reproduce the coding gains under like-for-like harnesses and effort settings; and whether any performance advantage survives after token use, accelerators and engineering overhead are counted together.
For now, K3 demonstrates that Moonshot can offer immediate access to a model with narrow frontier-level results. It does not yet demonstrate that the model is independently controllable, broadly superior or inexpensive to deploy. July 27 is when those claims can begin to be tested.
Meta reportedly plans to put departing AWS compute executive Dave Brown to work on its data-center buildout, adding hyperscale operating experience while leaving any customer-facing cloud business conditional and undefined.
China has paired a five-year AI training offer for developing countries with cooperation centers, a weather-warning rollout and a new 29-country organization. The package gives Beijing a platform for influence, but no budget, selection rules or delivery timetable has been published.
RoboTTT extends a robot policy’s visuomotor context to 8,000 timesteps and raises its average rubric score from 42% to 79%, but the evidence comes from three author-run assembly tasks and the longest task was completed in only two of ten trials.
Zhipu reportedly reached $1 billion in annual recurring revenue in July, roughly four times a March estimate, but the unconfirmed run rate is not annual sales and still sits far ahead of recognized cloud revenue while margins remain thin.
Nebius has arranged its first senior secured facility against an operating GPU deployment and one customer’s cash flows. The deal adds project-level debt to its expansion toolkit, but an unnamed customer and a larger rival financing limit what it proves about the rest of Nebius’s backlog.
Cursor says automatically running a repository-root git.exe on Windows does not meet its criteria for patching, while the researcher calls it an untrusted-search-path defect and separate research shows the same weakness across several competing AI coding tools.
Huawei publicly displayed a 16-cabinet Atlas 950 configuration rated at 1 EFLOPS in FP8, providing tangible evidence of its system-scale AI strategy while leaving price, power use and sustained workload performance undisclosed ahead of the full system's planned fourth-quarter release.
CIA Director John Ratcliffe said US intelligence is consistent with an estimate that Russian recruits last 20 to 30 minutes on Ukraine’s battlefield, but the public trail leads to an unsourced claim about assault troops and does not establish a representative average.
UK testing places leading open-weight models four to seven months behind selected closed-model cyber results, yet longer attack chains, U.S. benchmarks and mixed cost comparisons show why that interval is a warning signal rather than a universal capability clock.
TSMC reached the top of its second-quarter revenue guidance and raised its 2026 outlook, but a one-off investment gain boosted profit growth while the company committed more capital to 2-nanometer production, advanced packaging and an undated Arizona expansion.
Moonshot AI's Kimi K3 added to a global technology selloff with near-frontier performance, but unreleased weights, mixed cost comparisons and a recommended 64-accelerator deployment leave its effect on chip demand unresolved.
Two binding EU decisions require Google to give rival AI services comparable access to 11 Android features and offer eligible search competitors a restricted, anonymized dataset, but phased deadlines, certification, pricing and privacy safeguards leave the competitive effect unproven.
Databricks has signed a term sheet for a Coatue-led financing at a $188 billion valuation, while unidentified sources put the round at $3 billion. The proposed capital would deepen its push into AI governance, data agents and operational databases, but the transaction remains open and the company supplied no new operating figures.
Netflix says generative AI workflows were used on roughly 300 titles in 2026, but its only quantified example tied to that disclosure covers 17 minutes and does not establish how much finished material, spending or labor changed across the slate.
Gemini 3.5 Pro missed its expected June rollout. An anonymous-source account says a late-June data change fell short of Google's coding goals, but Google has confirmed only partner testing—not the reported cause, a new date, or public results and pricing.
Google is bringing Gemini Omni editing and a reusable face-and-voice avatar to Vids, but the sharper distinction is account-level identity across Vids and Gemini rather than a new category of AI video; Vids already offered Veo generation and customizable avatars, while specialist rivals already sell digital presenters.
Intel plans to adopt Gemini Enterprise across engineering, supply-chain and corporate work while extending chip-simulation capacity into Google Cloud, but the companies supplied no rollout schedule, contract value or performance baseline.
The Navy has approved an immediate department-wide framework for turning data into operational effects, but the public rollout leaves budgets, owners, deadlines, performance measures and assurance rules to a promised implementation roadmap.
Runta has raised a $20 million seed round led by Andreessen Horowitz for an execution layer that governs agents at the operating-system and network levels, entering a market where cloud runtimes and sandbox providers already offer overlapping controls.