Moonshot AI posts the full weights of Kimi K3 tomorrow at 00:00 UTC, the largest open-weight release yet. Google's Gemini 3.5 Pro is now 68 days past Sundar Pichai's May 19 promise with no date and no spec sheet.
Moonshot AI releases the complete weights of a 2.8 trillion parameter model at 00:00 UTC on July 27, described as the largest open-weight drop ever. Coverage this week credits K3 with beating Claude Opus 4.8 and GPT-5.5 on several coding and agent benchmarks while trailing Claude Fable 5 and GPT-5.6 Sol overall.
Ukraine's president said on July 25 that Moscow is seeking 30,000 North Korean soldiers, a sharp step up from earlier deployments already visible at May 2026's Moscow parade. The same day, Kazakh President Tokayev stood beside Putin and called for a freeze of the war.
Today's tool set targets the unglamorous layer: project structure, inbox triage, and calendar defense. Pair them with a self-hosted K3 endpoint and the automation stack stops depending on a single vendor's roadmap.
A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.
Join the founding list →Moonshot AI will release the full weights of Kimi K3, a 2.8 trillion parameter model the company describes as the largest open-weight release ever, this week at 00:00 UTC. Tech Times frames self-hosting as the point: locally run weights keep prompts off Moonshot's servers, which an API can never guarantee. K3 was...
Continue reading arrow_outwardMoonshot AI releases the full weights of Kimi K3, a 2.8 trillion parameter model it calls the largest open-weight release ever, on July 27 at 00:00 UTC.
The US paused strikes on Iran on Friday night after 13 consecutive nights and Tehran held fire through Saturday, but the Houthis said they hit Saudi sites on the Red Sea coast, widening the conflict as the central front quieted.
Mood is Fear: an open-weight Chinese model beating Claude Opus 4.8 and GPT-5.5 on coding benchmarks lands while a paused US-Iran air campaign reopens on a new front.
Reuters Breakingviews puts the tension plainly: a model that can be downloaded and self-hosted cannot be throttled at the API the way strikes can be paused or resumed, which is exactly why Washington is anxious about Kimi K3.
Reuters Breakingviews argues on July 25 that Kimi K3's distribution model, not its raw score, is the Washington problem: it can be downloaded and customized rather than gated behind an API. Tech Times makes the buyer-side version of the same argument, noting that locally run weights keep prompts off Moonshot's servers in a way no API can guarantee. Once the July 27 release lands, the usual leverage points over a frontier model disappear.
Coverage this week credits Kimi K3 with beating Claude Opus 4.8 and GPT-5.5 on several coding and agent benchmarks while still trailing Claude Fable 5 and GPT-5.6 Sol on overall capability. Read alongside the Frontend Code Arena result, the strength looks concentrated in frontend and agentic coding, not across the board. That is still enough to give buyers an open-weight option in the one workload they pay the most for.
Gemini 3.5 Pro is now 68 days past Sundar Pichai's May 19 Google I/O request to give the company until the following month, with no release date, no Gemini API entry, and no specification sheet. A leaked July 17 target also passed, and on July 21 Google shipped three cheaper, lighter Gemini models instead. The contrast with a 2.8 trillion parameter open-weight release timed to the minute is not flattering.
MCP Just Deleted Its Own Handshake
Read full analysis arrow_outwardMoonshot AI will release the full weights of Kimi K3, a 2.8 trillion parameter model the company describes as the largest open-weight release ever, this week at 00:00 UTC. Tech Times frames self-hosting as the point: locally run weights...
A Reuters Breakingviews roundup published July 25 argues Kimi K3's open-weight distribution is what makes it a policy problem in Washington: the model can be downloaded and customized rather than gated behind an API. Reuters says K3's...
Coverage this week credits Kimi K3 with beating Claude Opus 4.8 and GPT-5.5 on several coding and agent benchmarks, while still trailing Claude Fable 5 and GPT-5.6 Sol on overall capability. Read together with the Frontend Code Arena...
As of July 25, 2026, Google's flagship Gemini 3.5 Pro has no public release date, no Gemini API entry, and no specification sheet, 67 days after Sundar Pichai told Google I/O on May 19 to give the company until the following month. A...
The Model Context Protocol project published a July 28, 2026 specification release candidate that makes the protocol stateless: the handshake and session are gone, and server-to-client requests are restructured to be routable,...
Uber founder Travis Kalanick has raised $1.7 billion for Atoms, with Andreessen Horowitz and Uber joining the round, according to the Financial Times. Kalanick said the money supports work to automate industrial processes by deploying...
Ukrainian President Volodymyr Zelensky said on July 25 that Russia wants to bring 30,000 North Korean soldiers into the war, a sharp increase over Pyongyang's earlier deployments. North Korean servicemen took part in Moscow's military...
Speaking alongside Vladimir Putin at a bilateral forum on July 25, Kazakh President Kassym-Jomart Tokayev called for a freeze of the war in Ukraine, an unusually direct appeal from a leader inside Russia's traditional sphere of...
Deep Dive
7.2/ 10
Taskade stopped being a task app and became an app factory; the interesting question is whether a small team actually needs that.
This is for small teams and solo operators who want agents doing recurring work inside the same place the work is tracked, not in a separate automation tool. At $10/mo billed annually for unlimited agents and automations plus 50,000 AI credits, it is a cheap way to test agentic workflows without standing up a framework. If you mainly want a clean task board with solid reporting and permissions, the last two years of Taskade's roadmap have been moving away from you.
“Praise on Product Hunt centers on having task management, docs, collaboration, and AI agents in one workspace, with project structure generated straight from a prompt.”Product Hunt
Fyxer sits on Gmail or Outlook to triage the inbox, draft replies in your own tone, and write meeting notes automatically. The highest-value setup is to let it sort and pre-draft overnight, then spend ten minutes each morning approving or editing rather than composing from scratch. Treat the tone-matching as a first draft, not a send-and-forget.
Try it arrow_outwardReclaim automates your calendar by defending focus blocks and scheduling tasks and habits around meetings instead of letting meetings eat everything. Define two or three non-negotiable deep-work windows first, then let it shuffle lower-priority commitments around them. The discipline is in not overriding the blocks manually the moment someone asks for a call.
Try it arrow_outwardMotion builds a time-blocked schedule from your tasks, priorities, deadlines, and dependencies, rebuilding the day when something slips. Load real deadlines and effort estimates rather than vague to-dos, because the scheduling quality depends entirely on the input estimates. Useful for anyone juggling many small deliverables with hard dates.
Try it arrow_outwardThis free browser extension replaces the new tab page with a different well-designed 404 page each time, and it ships with credits and source code for each example. It is a low-effort way for designers and front-end developers to build a passive reference library of microcopy and error-state design. Use it as a swipe file: when you hit one you like, follow the source link and read the implementation.
Try it arrow_outwardStage your K3 download before 00:00 UTC. Provision storage and a serving runtime tonight so tomorrow's 2.8 trillion parameter release lands into a working harness instead of a scramble. Pick one coding task you already run on Opus 4.8 and keep it as the comparison case.
Build a two-model routing rule. Send coding and agent work to Kimi K3 where this week's benchmarks favor it, and keep general reasoning on GPT-5.6 Sol or Claude Fable 5 where K3 still trails. Log cost per completed task, not per token.
Reclaim one hour with Fyxer and Reclaim. Wire Fyxer onto your inbox for triage and let Reclaim defend two focus blocks this week, then track the tasks inside Taskade so the saved hour has a destination.
Reuters Breakingviews made the sharpest point of the week on July 25: Kimi K3's problem for Washington is not its score, it is that anyone can download and customize it. Meanwhile Google shipped three cheaper Gemini models on July 21 and still cannot name a day for 3.5 Pro, 67 days after Pichai asked for one more month. For builders, that asymmetry is the whole story - one lab hands you the artifact, the other hands you a waiting list. Spend today making your stack able to accept the artifact.
AI analysis, world news, markets, and tools. One briefing, delivered free.