Moonshot's Kimi K3 is the first open 3-trillion-class model, and it moved markets on release day

Good morning 👋 A Chinese lab just shipped a model that only two labs on earth beat, priced at a fraction of what those two charge, and the market reacted before most of us finished coffee.

In today's issue:

• 🔭 Moonshot's Kimi K3 rivals the frontier at open-weight prices, and markets flinched
• 🧠 Nemotron 3 Embed's retrieval win, Grok lands on Bedrock, Google Search gets hands
• 🔬 Robots with 8,000-step memory, comment-section data poisoning, and what security agents actually cost
• 🛠️ How to stress-test Kimi K3 against your own prompts before you trust it
• 🔗 ChatGPT's teen guardrails, Cars24's voice-agent numbers, a biosecurity framework worth a skim

🔭 THE ONE THING

🌊 Moonshot's Kimi K3 is the first open 3-trillion-class model, and it moved markets on release day

Moonshot AI shipped Kimi K3: 2.8 trillion parameters, a 1M-token context window, and a claim to be the world's first open model in the 3T class. Weights land by July 27. The efficiency story matters more than the size flex: Moonshot's new Delta Attention plus Attention Residuals combo gets roughly 2.5x the scaling efficiency of Kimi K2, and Artificial Analysis independently clocks it behind only Claude Fable 5 and GPT-5.6 Sol on their Intelligence Index. That's a real ranking, not Moonshot's own benchmark deck. At $3/MTok input and $15/MTok output, it undercuts the frontier labs by a wide margin for a model sitting third in the world and open weight to boot.

🧠 MODELS & RELEASES

• 🥇 NVIDIA Nemotron 3 Embed took the #1 spot on RTEB, the retrieval benchmark built to catch what actually breaks agents: missed context, wasted re-queries, noisy reasoning chains. 78.5% overall score, open weights, commercially licensed, with a 1B variant tuned for Blackwell GPUs. Worth benchmarking against whatever's powering your RAG stack right now. link
• ⚡ xAI Grok 4.3 is live on Amazon Bedrock: tool calling, structured output, image input, dialable reasoning effort. The model isn't new. The distribution is. Any enterprise already locked into AWS contracts and compliance sign-off can now call Grok without touching xAI's own API. link
• 🔍 Google AI Mode can now link into your actual third-party apps and act inside them, not just answer questions from a distance. Small feature note, big signal about where Google wants Search traffic to go next. link

🔬 RESEARCH HIGHLIGHTS

• A robot policy that remembers 8,000 steps back, at no extra inference cost. Most robot foundation models look at the last frame or two and nothing else. RoboTTT scales working memory to 8,000 timesteps, three orders of magnitude past prior work, by folding test-time training into the policy instead of just widening the context window. The payoff: one-shot imitation from watching a single human demo, on-the-fly correction after a perturbation, and a five-minute, ten-stage assembly task that no shorter-context baseline could finish at all. Memory length, not just model size, looks like the next real scaling knob for embodied AI. arxiv.org/abs/2607.15275

• Your pretraining corpus can be poisoned from the comments section. Earlier poisoning research stuck to tidy, well-behaved sources like Wikipedia. This paper points at something messier and already public: open discussion interfaces, the comment threads and forums that get swept into web crawls. The harder question is whether injected content actually survives data curation and lands in a trained model, which is why the authors built a measurement tool (HalfLife) to check. No headline attack-success number here, and that's the honest state of the research, but the vector itself is the finding: if your data pipeline crawls the open web, someone else's comment thread is part of your training set. arxiv.org/abs/2607.15267

• A security-agent benchmark that finally asks what the tool calls cost. Most agent security evals just measure who wins, assuming an unlimited compute budget, which is not how a SOC operates. This one scores agents on both offense (Cybench) and defense (Splunk BOTS v1 investigations) against actual inference and tool spend. The split is the interesting part: offensive capability climbs cleanly with more test-time compute, and cheaper open-weight models can close the gap on proprietary ones. Defensive work doesn't scale that way at all — investigating an incident rewards disciplined tool use and knowing what to query, not throwing more reasoning budget at the problem. Worth remembering before anyone sells you a bigger model as the fix for a slow SOC. arxiv.org/abs/2607.15263

🛠️ TRY THIS

Give Kimi K3 a real tryout before you trust it in production

1. Point your current agent harness at Kimi K3's endpoint alongside your primary model, no swap yet, just parallel calls.

2. Run last week's actual failed or edge-case prompts through both, not a public benchmark.

3. Diff the five or ten outputs that matter most to your product. Ignore the ones that were fine either way.

4. If parity holds on your own prompts, make the swap a config change, not a rewrite (see Omnigent below).

Prompt: Compare these two model outputs on the same input. Flag any factual divergence, then say which one you'd ship to a user without editing.

Worth a look

• Omnigent is an open-source meta-harness for orchestrating Claude Code, Codex, Cursor, and custom agents under one policy and sandboxing layer, so adding Kimi K3 as an option means a config change instead of a rewrite. omnigent-ai/omnigent
• PixelRAG skips HTML parsing entirely and treats a page as pixels for retrieval. Early, but if your RAG pipeline keeps breaking on site redesigns, it's worth a look. StarTrail-org/PixelRAG

🔗 QUICK LINKS

• OpenAI rolled out age-gating and parental controls for ChatGPT, betting it can make teen safety a feature instead of a lawsuit magnet. link
• Cars24 runs 1M+ monthly conversation minutes through OpenAI voice and chat agents and claims a 12% lift in recovered leads. Vendor case study, but the volume number is real. link
• Google DeepMind and Isomorphic Labs published a joint bioresilience framework for AI models. Worth a skim if you're anywhere near biosecurity policy, otherwise skip. link
• Built Technologies built a document-intelligence engine on AWS (with the GenAI Innovation Center) that classifies and extracts across real estate finance paperwork. Narrow use case, but a clean template for anyone drowning in PDFs. link

That's today's issue. Back tomorrow with whatever Kimi's weights actually reveal on July 27.

Pradeep Perugu