Sign in Subscribe
AI Intelligence

Moonshot Unlocks the First Open 3T-Class Model

Vijay Bhagwati

• 8 min read
Share

Moonshot's Kimi K3 joins the frontier tier as an open-weight 2.8-trillion-parameter model, Microsoft publicly questions Anthropic's Fable guardrails, and European hardware startups raise at record seed checks.

July 18, 2026 | Reading time: 11 minutes | Issue #214

Lead

Beijing-based Moonshot AI released Kimi K3 on Thursday, billing it as the first open 3-trillion-class model and the largest open-weight release to date at 2.8 trillion parameters. According to the company's technical blog, the model uses a mixture-of-experts architecture with 16 of 896 experts active per token, Kimi Delta Attention, and Attention Residuals, and it carries a 1-million-token context window with native vision. Moonshot says K3 reaches frontier-level performance on long-horizon coding and reasoning benchmarks and will publish full weights by July 27, 2026. Independent evaluation by Artificial Analysis on the AA-Briefcase long-horizon knowledge-work benchmark placed K3 second only to Claude Fable 5 Max, ahead of GPT-5.6 Sol Max, while Moonshot's own suite shows K3 leading on BrowseComp and Automation Bench (VentureBeat).

The release lands two days before the World Artificial Intelligence Conference in Shanghai and immediately reshapes the open-source landscape. Chinese labs now hold the two largest open-weight releases — DeepSeek V4 Pro at roughly 1.6 trillion parameters and Kimi K3 at 2.8 trillion — and both are being measured against the most capable closed systems from OpenAI and Anthropic. For Moonshot, K3 is a comeback play: after DeepSeek's R1 disruption in early 2025 eroded Kimi's user rank in China from third to seventh, the company pivoted toward open releases starting with Kimi K2. The new model's API pricing is set at $0.30 per million cache-hit input tokens, $3 for cache-miss input, and $15 for output, with compatibility for the OpenAI SDK.

K3 also arrives with a concrete agentic demo. Moonshot says the model designed, optimized, and verified a 4 mm² chip in a 48-hour autonomous run using open-source EDA tools, achieving timing closure at 100 MHz and simulated decode throughput above 8,700 tokens per second. The claim is not a production roadmap, but it signals where Moonshot thinks the next battleground sits: not benchmark comparisons alone, but models that can sustain long-range technical work with minimal human oversight.

Briefs

OpenAI Trains a Red-Teaming Model to Harden GPT-5.6

OpenAI published a detailed account of GPT-Red, an automated red-teaming model trained to find prompt-injection vulnerabilities and adversarially harden GPT-5.6. The company says GPT-Red breaks nearly all internal and production models up to GPT-5.5 and achieves an 84% success rate on an indirect prompt-injection benchmark compared with 13% for human red-teamers. GPT-5.6 Sol is now the most robust production model to direct prompt injections OpenAI has shipped, with six times fewer failures than its best model from four months earlier. OpenAI keeps GPT-Red internal, separating the attacker from the deployed model. The work is the clearest operationalization yet of the lab's argument that safety must scale alongside capability, and that automated red-teaming is one path to self-improvement for robustness (OpenAI).

Microsoft Unifies Copilot and Singles Out Anthropic's Fable Controls

Microsoft CEO Satya Nadella told engineers this week that Anthropic's request limits on its top-tier Fable model "don't make sense," CNBC reported. In remarks provided to the network, Nadella said: "If you use Fable, when it refuses for any random thing, it just is like, when was the last time you had a creation tool that was so editorially controlled?" The comments come as Anthropic routes some user queries on large-model construction to older model versions and has faced social-media complaints about refusals. Nadella also used the meeting to argue that Microsoft's consumer and corporate Copilot experiences should be unified and that "it can't be that there are only two companies in the world with token capital, and everybody else is renting it." Microsoft has a $5 billion stake in Anthropic and is a major Azure customer for the startup (CNBC).

Google Renames NotebookLM and Pushes AI Biology as a Security Tool

Google rebranded NotebookLM as Gemini Notebook, embedding it more deeply into the Gemini app and Google Search and adding a secure cloud-computer feature that can write and execute code inside notebooks. The code-execution capability is live first for AI Ultra users and Workspace business customers, with wider Pro rollout planned. Separately, Google DeepMind announced a bioresilience program with Isomorphic Labs aimed at using frontier AI for pathogen surveillance, vaccine design, and outbreak response. The framing is explicit: the same models that could amplify biological risk can be redirected to prevent it. DeepMind VP Helen King told Axios the lab would not launch a model if it crossed a critical capability threshold without appropriate mitigations, though she said that threshold has not been reached (Google, Axios).

Meta Adds Parental Alerts for Teen AI Conversations

Meta said Thursday it will notify parents when a teen discusses suicide or self-harm with Meta AI, and is building the ability to contact emergency services when conversations indicate imminent risk. The alerts, live for Instagram Parental Supervision users in the U.S., U.K., Australia, and Canada, are manually reviewed before being sent and will roll out globally by year-end. Meta is also extending its "Limited Content" setting for teens to Meta AI. The move comes amid broader regulatory and legal pressure on how AI chatbots respond to minors in crisis (TechCrunch).

From the Lab

The latest arXiv drop in cs.AI, cs.LG, and cs.CL includes several papers that move beyond standard benchmark optimization into the operational costs of deploying models and agents. "Beyond Success Rate: Cost-Aware Evaluation of Offensive and Defensive Security Agents" by Kassianik, Nelson, and Singer introduces a framework for measuring the cost, not just the accuracy, of autonomous security agents — a dimension that becomes central as labs release models with explicit cyber capabilities. AISI's public analysis this week that leading open-weight cyber models now trail the closed frontier by four to seven months, down from six to ten months in 2025, makes this cost-aware lens more urgent: the gap is closing, and defenders will need to know the price of each capability step.

"Bridge Evidence: Static Retrieval Utility Does Not Predict Causal Utility in Multi-Step Agentic Search" by Mukhopadhyay et al. takes on the assumption that a document useful in isolation will remain useful inside a longer reasoning chain. The authors argue that standard retrieval metrics mislead when documents are used for causal reasoning across multiple agent steps. The finding has practical implications for products like Gemini Notebook, Kimi Work, and OpenAI's ChatGPT Work, all of which promise to synthesize large document sets over long sessions. If retrieval quality decays as agents chain steps, the bottleneck may shift from model scale to how agents track and re-query sources. A third paper, "Pretraining Data Can Be Poisoned through Computational Propaganda" by Graf et al., shows how coordinated, low-cost campaigns can introduce behaviors into language models during pretraining that are hard to detect afterward. The result adds another reason for frontier labs to treat data provenance as a first-class safety concern alongside red-teaming and capability evaluation (arXiv).

Eastern Front

Kimi K3 is the headline, but the broader Chinese AI economy is reorganizing around tokens as an internal currency. A South China Morning Post report published Friday described how ByteDance employees now burn through close to a billion AI tokens per month as part of their work, with Alibaba, Baidu, and Tencent building similar internal token-accounting systems. The piece frames the falling cost of inference as turning compute into a corporate budgeting line item rather than a capital project. The same dynamic helps explain why Chinese labs can release enormous open-weight models at prices that undercut Western APIs: the marginal cost of a token is becoming a known, optimizable quantity, and competition is pushing it down faster than expected.

The geopolitical layer remains active. A Chinese filing this week implied a DeepSeek valuation near $52 billion, Reuters reported, while the lab continues to prepare for an IPO that could value it at $71 billion. Washington is responding by widening scrutiny of U.S. corporate use of Chinese models and, according to Axios, the Trump administration has twice intervened this summer over model access. The policy argument is no longer theoretical: U.S. companies are adopting cheaper Chinese systems, and lawmakers are trying to decide whether to block, audit, or match them.

India Lens

There is no major India AI story in Friday's river, but the regional signal is structural. India remains the largest open-source model adopter by API volume on platforms like OpenRouter and Fireworks, which are precisely the distribution channels Chinese labs are targeting with Kimi K3 and DeepSeek. The country's developers and enterprises are price-sensitive, data-localization requirements are tightening, and the government has been explicit about preferring domestic or auditable infrastructure. That positions India as a proving ground for whether open-weight Chinese models can convert benchmark attention into enterprise share outside China. The absence of a funding or product headline today does not mean the market is quiet; it means the competition there is now downstream of the model release, fought over inference pricing, hosting partnerships, and compliance certifications rather than splashy launches.

The View

The dominant pattern this week is fragmentation at the frontier. A year ago, the top tier looked like a closed oligopoly: OpenAI and Anthropic set the benchmark pace, Google chased, and everyone else picked up crumbs. Now Moonshot has released a 2.8-trillion-parameter open model that independent evaluators place near the top on long-horizon knowledge work, Microsoft is publicly criticizing Anthropic's safety controls while deepening its own model pluralism, and capital is organizing around inference chips as collateral. The competitive unit is no longer the model; it is the stack — model, compute, distribution, and trust architecture — and the stack is being unbundled.

For U.S. labs, the risk is that their safety and export-control maneuvers become a tax their open-source competitors do not pay. Anthropic's Fable restrictions, however principled, are being weaponized in competitive messaging by a partner that has also invested in the company. OpenAI's GPT-Red is a strong technical answer to the robustness problem, but it does not address the pricing pressure from models that can be self-hosted. The next phase of the AI market will be defined by who can deliver frontier capability at a cost and control profile that enterprises actually want, not just by who wins the next leaderboard.

The Miss

Nvidia's Cosmos 3 Edge launch and Japan physical-AI expansion got headline attention, but the more consequential detail may be the company's push into healthcare and biotech through the Tokyo-1 drug-discovery consortium and partnerships with Astellas, Daiichi Sankyo, and Ono Pharmaceutical. The market still treats Nvidia as a chip company, yet its repeated moves into vertical AI platforms — BioNeMo, drug-discovery agents, industrial robotics — suggest it is building application-layer leverage long before the hardware cycle peaks. The chip narrative is louder; the platform narrative is more durable.

Pull Quotes

"If you use Fable, when it refuses for any random thing, it just is like, when was the last time you had a creation tool that was so editorially controlled?" — Satya Nadella, Microsoft CEO, in remarks reported by CNBC

"It can't be that there are only two companies in the world with token capital, and everybody else is renting it." — Satya Nadella, Microsoft CEO, via CNBC

"The next frontier of AI is in the physical world, and this is a once-in-a-generation opportunity for Japan." — Jensen Huang, Nvidia CEO, via CNBC

"Open-source is no longer lagging six months behind Western closed-source models." — AI commentator Kimmonismus, quoted by VentureBeat

The open-weight tier is now too close to ignore — and too big to block.