Moonshot's Kimi K3 is now state television's proudest export, with weights, a technical report, and three training methods published. Three of today's five AI stories are Chinese releases.
Kimi K3 landed at 2.8 trillion parameters with weights, technical report, and three training methods published, and CCTV gave it a Focus Interview segment. DeepSeek shipped its first multimodal vision model and SenseTime open-sourced SenseNova U1.5 Lite the same day.
AI Mode is now sometimes served by Gemini 3.7 Flash with no benchmark data attached, per reporting tied to Google's August 2026 spam update. A free year of Gemini for students at semester start buys surface area rather than scores.
Russian Drones Kill 15 At Ukraine Mall: a busy shopping centre in Kryvyi Rih was struck on 21 August, with more than 120 wounded and a second drone hitting the roof about 30 minutes later. The ICC president warned the same day that US sanctions threaten the court itself.
Koda publishes a fact-checked AI intelligence briefing every morning: site, email, podcast, and video. Founding sponsors lock today's rates for 6 months as the audience compounds.
View the media kit →DeepSeek launched DeepSeek-V4-Flash-Vision-Exp, an experimental visual-understanding model, and opened its multimodal API service to developers. The company says it matches DeepSeek-V4-Flash on text tasks including agents, reasoning, and world knowledge while adding image understanding on the same API endpoint developers already use...
Continue reading arrow_outwardDeepSeek shipped DeepSeek-V4-Flash-Vision-Exp and opened a multimodal API, its first vision model inside the V4 API line, claiming parity with V4-Flash on agents, reasoning, and world knowledge.
Russian drones struck a shopping centre in Kryvyi Rih on 21 August, killing 15 and wounding more than 120, with a second drone hitting the roof roughly 30 minutes later in a pattern Zelensky said targeted emergency responders.
Mood reads Greed, with Chinese labs releasing frontier weights for free while Google and OpenAI spend on distribution rather than defending price.
The ICC's sanctioned officials are losing banking, software, and cloud access at the same moment SenseTime and Moonshot publish weights anyone can run locally: sanctions bite services, not files.
Three of today's five AI stories are Chinese releases: Kimi K3 at 2.8 trillion parameters with weights, technical report, and three training methods published; DeepSeek's first vision model; SenseTime's SenseNova U1.5 Lite. CCTV giving Moonshot a Focus Interview segment signals this is now industrial policy with a broadcast arm, not just lab competition.
Google routed AI Mode to Gemini 3.7 Flash with no benchmark data attached and handed students a free year of Gemini at semester start. Both moves buy surface area and habit, and SenseTime's pitch for U1.5 Lite similarly leads on practical image editing usability rather than leaderboard position.
Trump's 21 August window admits up to 300,000 metric tons of ground beef over 90 days free of out-of-quota tariffs. The administration is willing to puncture its own trade architecture the moment food prices become a political liability, which tells exporters where the real pressure points sit.
200 Million Miles, Zero Tornadoes
Read full analysis arrow_outwardVercel shipped fx, a tiny open-source coding agent aimed at fast edits and prototype generation rather than repo-wide refactors. The value of a small agent is auditability: read the source, see exactly what it does to your filesystem, then decide how much rope to give it. Use it for throwaway scaffolding and compare its output against whatever heavier agent you already pay for.
Try it arrow_outwardGoogle's Antigravity agents now run as extensions inside an existing editor instead of forcing you into a separate agent surface. That removes the main friction of agentic coding, which is context switching between the tool that writes code and the tool where you review it. Test it on one branch with a narrow task, then check whether the diffs are small enough to review honestly.
Try it arrow_outwardLaunched August 20 to 21 and reported as free, Task Engine 2.0 embeds an office agent directly in Baidu search so a query can carry through to office work, deep research, and even app generation. The interesting design lesson is packaging: the agent lives where the information hunt already starts, not in a separate console. Worth a scouting pass if you are designing research-to-deliverable workflows.
Try it arrow_outwardThe Iris.ai Researcher Workspace is built to analyze and summarize bodies of research data rather than answer one-off questions, which suits literature reviews and technical due diligence. Give it a corpus you already know well and check whether its summaries survive a spot-check against the source papers. That test tells you more than any benchmark claim on the marketing page.
Try it arrow_outwardAI Reality generates simple web AR prototypes using Stable Diffusion and OpenAI models, so you can put a spatial concept in front of stakeholders before committing engineering time. Treat the output as a pitch artifact rather than a product: it is for testing whether the idea reads at all. If it does, you have earned the argument for a real build.
Try it arrow_outwardBenchmark the Chinese trio yourself. Pull the Kimi K3 technical report and its three published training methods, then run DeepSeek-V4-Flash-Vision-Exp and SenseNova U1.5 Lite against your own image editing and document tasks. Vendor claims of parity mean nothing until they meet your data.
Instrument your Gemini 3.7 Flash exposure. AI Mode routing changed with no benchmark disclosure, so log your own latency and answer quality on search-dependent flows this week. If your traffic or citations shifted after the August 2026 spam update, you want that baseline before the next silent swap.
Ship one agent into the editor you already use. Read fx by Vercel end to end before running it, wire Antigravity IDE Extensions into your existing setup, and try Task Engine 2.0, free since its 20 to 21 August launch, on a task that must end in a finished document.
Moonshot gave away 2.8 trillion parameters and got state television coverage for it; Google gave away a year of Gemini and got the semester. Neither move came with benchmark data, because neither is competing on benchmarks anymore. For builders that means capability is arriving cheaper than your procurement cycle can absorb, and the moat is now whatever you wrap around it. Test three open models this week and see how much of your stack is actually differentiated.
AI analysis, world news, markets, and tools. One briefing, delivered free.