# Kimi K3 Open Weights Narrow Frontier Gap as NVIDIA Bets on Open Security *By AI High Signal Digest • July 28, 2026* Moonshot AI's 2.8T-parameter Kimi K3 becomes the first open-weight model to approach frontier performance, narrowing the open-closed gap to 4 points. NVIDIA launched an open security alliance and invested in SSI, while Anthropic staked out its open-weights position and Microsoft entered cybersecurity AI. ## Top Stories *Why it matters: the open-weight frontier crossed a new threshold, while NVIDIA consolidated its position as the infrastructure backbone for both open security and frontier safety labs.* **Moonshot AI released Kimi K3 open weights—the largest open-weight model to date and the first to approach frontier performance.** The 2.8T-parameter MoE model (104B active, 16 of 896 experts) features native multimodal support, a 1M-token context window, and MXFP4 weights via quantization-aware training. [^1] K3 scores 57 on the Artificial Analysis Intelligence Index; only Anthropic's Opus 5 (61), Fable 5 (60), and OpenAI's GPT-5.6 Sol (59) score higher, narrowing the open-closed gap to 4 points—the smallest since February. [^2] K3 (Max) ranks #1 among open-weight models in Agent Arena (+9.75% net improvement) and #1 overall in Frontend Code Arena. [^3][^4] The model launched with day-0 support from Together, Fireworks, Nebius, vLLM, SGLang, Cursor, and Devin, with third-party pricing at $5.40 blended per million tokens—matching Kimi's direct pricing. [^5][^6] The modified-MIT license requires model-as-a-service businesses over $20M revenue to secure a separate commercial agreement. [^7] **NVIDIA launched the Open Secure AI Alliance and invested in Safe Superintelligence, expanding its footprint across AI security and safety.** Jensen Huang argued defenders need an open ecosystem because attackers already have frontier AI, citing the Hugging Face incident where closed AI blocked forensics while an open-weight model helped contain the breach. [^8] Andrew Ng endorsed the move, calling claims that closed models are safer "just regulatory capture." [^9] Separately, NVIDIA's substantial SSI investment enables a 10x compute increase and access to Vera Rubin GPUs; SSI had previously relied on Google TPUs. [^10][^11] ## Research & Innovation *Why it matters: AI crossed another mathematical frontier, and Kimi K3's technical report revealed numerical stability as the defining challenge at 3T scale.* **AI solved a second FrontierMath: Open Problems benchmark.** A presentation for the absolute Galois group of the 2-adic numbers—open for over 40 years—was produced in collaboration with David Roe, the problem's proposer. [^12][^13] **Kimi K3's technical report centers on numerical stability at 3T scale.** The architecture mixes Kimi Delta Attention and Gated MLA in a 3:1 ratio with qk-norm constraints to manage dynamic range. [^14][^15] The report does not disclose training tokens; observers estimate 3–4×10²⁵ FLOPs, making it the largest model trained in China. [^16] ## Products & Launches *Why it matters: cybersecurity AI moved from concept to production-grade deployment at lower cost.* **Microsoft released MAI-Cyber-1-Flash, its first cybersecurity model.** Combined with the MDASH multi-agent harness, it scores 96% on CyberGym—12 points above Anthropic's Mythos—at half the cost, handling up to 90% of vulnerability detection and patching tasks. [^17][^18] ## Industry Moves *Why it matters: compute financing and semiconductor sovereignty are reshaping the competitive landscape.* **NVIDIA is reportedly considering guaranteeing up to $250B** to back OpenAI's lease on a $500B data center, plus separate financing discussions for up to $350B in chip purchases. [^19] **China began mass-producing domestic DUV lithography tools**, a state-backed advance in reducing foreign reliance; ASML shares fell 6% on the news. [^20][^21] ## Policy & Regulation *Why it matters: open-weights policy and pre-release government review converged, with Anthropic staking out a distinctive position.* **Anthropic published its open-weights position.** Dario Amodei wrote the company "has never advocated for a ban" but wants mandatory safety testing for all sufficiently capable models, chips kept out of China, and industrial-scale distillation stopped. Anthropic did not sign the NVIDIA-led coalition letter. [^22][^23] **A US framework nearing completion** would give federal agencies up to 30 days of pre-release access to frontier models; OpenAI, Anthropic, and Google are negotiating the rules. [^24] Sam Altman heads to Washington this week to preview OpenAI's newest model. [^25] ## Quick Takes - **Opus 5 drew sharp user backlash** for reliability issues—endless loops, unfinished tasks—with some developers switching back to Fable, despite Opus 5 Max ranking #1 in Frontend Code and Text Arenas. [^26][^27][^28] - **Cognition's Devin** now writes ~95% of Cognition's code; the company hit a $500M+ run rate, up from $37M last year. [^29] - **ChatGPT Sites** now hosts over one million sites. [^30] --- ### Sources [^1]: [𝕏 post by @mervenoyann](https://x.com/mervenoyann/status/2081762511276110315) [^2]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2081926991788626011) [^3]: [𝕏 post by @arena](https://x.com/arena/status/2081804108433072623) [^4]: [𝕏 post by @arena](https://x.com/arena/status/2081809520184209518) [^5]: [𝕏 post by @jaminball](https://x.com/jaminball/status/2081829990015127743) [^6]: [𝕏 post by @cognition](https://x.com/cognition/status/2081766141454925992) [^7]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2081821449745236270) [^8]: [𝕏 post by @JensenHuang](https://x.com/JensenHuang/status/2081698060330250294) [^9]: [𝕏 post by @AndrewYNg](https://x.com/AndrewYNg/status/2081787106062746002) [^10]: [𝕏 post by @ssi](https://x.com/ssi/status/2081732119194394763) [^11]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2081740668125225229) [^12]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2081894720813604997) [^13]: [𝕏 post by @DavidTurturean](https://x.com/DavidTurturean/status/2081780318881677693) [^14]: [𝕏 post by @suchenzang](https://x.com/suchenzang/status/2081773594347274516) [^15]: [𝕏 post by @suchenzang](https://x.com/suchenzang/status/2081776318488973604) [^16]: [𝕏 post by @teortaxesTex](https://x.com/teortaxesTex/status/2081764014883848563) [^17]: [𝕏 post by @mustafasuleyman](https://x.com/mustafasuleyman/status/2081781833100820681) [^18]: [𝕏 post by @satyanadella](https://x.com/satyanadella/status/2081779755146482153) [^19]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2081677231353663727) [^20]: [𝕏 post by @jukan05](https://x.com/jukan05/status/2081727917885903032) [^21]: [𝕏 post by @amir](https://x.com/amir/status/2081832674738786689) [^22]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2081873374926540978) [^23]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2081874065908785158) [^24]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2081836187506065701) [^25]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2081870609613730132) [^26]: [𝕏 post by @RaoulGMI](https://x.com/RaoulGMI/status/2081807030826705146) [^27]: [𝕏 post by @Teknium](https://x.com/Teknium/status/2081896043202158930) [^28]: [𝕏 post by @arena](https://x.com/arena/status/2081831019377004727) [^29]: [𝕏 post by @MollySOShea](https://x.com/MollySOShea/status/2081749064056717807) [^30]: [𝕏 post by @JonathanAbrams](https://x.com/JonathanAbrams/status/2081819354367451186)