# Kimi K3 Debuts as AI Science Labs, World Models, and Inference Scale Up *By AI News Digest • July 17, 2026* Moonshot AI unveils Kimi K3 and plans to open its weights, while China reiterates its open-AI posture. The digest also examines Lila’s experiment-driven reasoning approach, Yann LeCun’s push for world models, and Fireworks AI’s major funding round. ## Open models and access policy ### Kimi K3 arrives with frontier-scale specifications and planned open weights Moonshot AI introduced **Kimi K3**, a native multimodal model with 2.8 trillion parameters and a one-million context window. The company says its Delta Attention enables up to 6.3× faster decoding at million-context scale, while Attention Residuals raise training efficiency by roughly 25% for under 2% additional cost; it is positioning the model for long-horizon agentic coding and self-evolving workflows. [^1] K3 is available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with open weights scheduled for July 27. Emad Mostaque separately estimated its training at about 1e25 FLOPs and $15–25 million, noting that the underlying training details have not all been disclosed. [^1][^2] *Why it matters:* Nathan Lambert characterized K3 as a large multimodal planning model for difficult tasks, alongside distinct open-model options for agentic work, general multimodality, and lower-cost serving. He called the current group of releases the most relevant open-model set since DeepSeek R1. [^3][^4] ### China reiterates an open-AI stance at the World AI Conference A summary of President Xi Jinping’s first appearance at Shanghai’s World AI Conference said he reaffirmed China’s commitment to open source and “openness and win-win” AI. The speech also warned against overextending national-security concepts in AI and pledged 5,000 AI training and seminar opportunities for developing countries over the next five years through groups including ASEAN, the African Union, and BRICS. [^5] *Why it matters:* The speech places international access and open-source development at the center of China’s stated AI posture. Nathan Lambert read it as a commitment to continuing an “open, global” approach. [^6] ## AI systems move into labs and the physical world ### Lila Sciences treats experiments as data for reasoning models Lila Sciences is building automated “AI Science Factories” in which models propose experiments, receive experimental feedback, and use nature as a verifier for reinforcement-learning post-training. The company says it has assembled 10 trillion experimentally verified scientific reasoning tokens across life sciences, chemistry, and materials science, and reports that its general model often outperforms domain-specific models. [^7] Its lab architecture connects instruments through a physical transport layer; instructions are API calls that may be executed by either robotic or human arms. Lila describes the lab platform as the data-generation mechanism, while the reasoning model is the core asset. [^7] *Why it matters:* This is a concrete attempt to extend verifiable-reward training beyond math and code by making experimental outcomes part of the learning loop. Lila’s thesis is that scientific experimentation can provide the next large-scale source of post-training data. [^7] ### Yann LeCun’s AMI Labs pursues world models for physical AI Yann LeCun described AMI Labs’ focus as world models that learn from physical signals and predict the consequences of actions. Its JEPA approach learns abstract representations from video and predicts in latent space rather than reproducing pixels; LeCun said V-JEPA models can already identify certain physically impossible events in video with limited fine-tuning. [^8] LeCun argued that reliable agentic systems, domestic robots, and level-5 self-driving require such models to reason about the real world. AMI is also advancing Project Tapestry, a distributed-training framework intended to let countries contribute local data and compute to shared foundation models without disclosing the raw data; its kickoff took place in Paris. [^8] *Why it matters:* The work targets a capability gap that language-model scaling alone does not address: planning and acting under physical constraints. ## Inference business reaches new scale ### Fireworks AI raises $1.5 billion as specialized-model usage grows Fireworks AI announced a **$1.5 billion Series D** at a **$17.5 billion valuation**. The company also said it has surpassed $1 billion in annual recurring revenue and serves more than 40 trillion tokens daily, with over 95% of those tokens coming from models specialized on customer data. [^9] *Why it matters:* The financing and operating figures point to a large market for production inference and customer-specific models. Nathan Lambert argued that inference companies with billion-dollar-scale revenue may become functioning “neolabs” faster than many research labs. [^10] --- ### Sources [^1]: [𝕏 post by @Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2077830229968683203) [^2]: [𝕏 post by @EMostaque](https://x.com/EMostaque/status/2077828058489209184) [^3]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2077777704573919598) [^4]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2077777436553720019) [^5]: [𝕏 post by @vince_chow1](https://x.com/vince_chow1/status/2077947375964791028) [^6]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2077979813776970007) [^7]: [🔬 RL with Verifiable Rewards, but the Verifier is a Lab — Lila Sciences](https://www.youtube.com/watch?v=2wIxPWK6nCs) [^8]: [Fireside Chat with Yann LeCun, Executive Chairman of AMI Labs | RAISE Summit 2026](https://www.youtube.com/watch?v=iDpPFAXmcZc) [^9]: [𝕏 post by @FireworksAI_HQ](https://x.com/FireworksAI_HQ/status/2077746648193941833) [^10]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2077754380842082528)