Models & Labs
Moonshot AI's Kimi K3 Model Bets on Memory
AI Newshigh confidence
Why it matters
- →Kimi K3 is the largest open-weight model, setting a new benchmark in AI model size.
- →The model's focus on memory efficiency over compute power addresses China's hardware constraints.
- →It offers a viable option for enterprises seeking data sovereignty and regional language support.
Moonshot AI has launched the Kimi K3, the largest open-weight AI model with 2.8 trillion parameters. The model emphasizes memory efficiency over compute power, using techniques like mixture-of-experts and quantization-aware training to reduce memory demands. This design choice is strategic, given China's limitations on high-performance computing resources. While K3 requires significant infrastructure to deploy, it offers Asian enterprises a robust option for data sovereignty and regional language support. The model's release marks a significant shift in AI model design, focusing on memory as a key resource.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1Kimi K2.6 Powers Autonomous Code Generation
Cole Medin · April 25, 2026 · Background
- 2Moonshot AI raises $2B at $20B valuation
TechCrunch AI · May 7, 2026 · Related
- 3JetBrains Releases Mellum2: Efficient 12B MoE Model
Hugging Face Blog · June 1, 2026 · Background
- 4Together AI Hosts MiniMax M3 for Efficient Inference
Together AI Blog · June 2, 2026 · Background
- 5Kimi K2.7 and GLM-5.2 Models Released
Lev Selector · June 19, 2026 · Same story
- 6OpenAI and Broadcom unveil LLM-optimized chip
OpenAI · June 24, 2026 · Background
- 7MiniCPM5: A 1B Model for On-Device Projects
Sam Witteveen · July 5, 2026 · Background
- 8Thinking Machines Lab Releases Inkling Model
WIRED AI · July 15, 2026 · Background
What happened next
- 9Moonshot's Kimi K3 Model Approaches Fable 5 Benchmarks
The AI Daily Brief · July 21, 2026 · Related
- 10Kimi K3: Largest Open-Weight AI Model Released
Together AI Blog · August 1, 2026 · Same story
- 11Moonshot AI Targets $2B Revenue with Open-Weight Model
TechCrunch AI · September 11, 2026 · Same story
- 12Moonshot AI releases Kimi K3.1 Pro
Matthew Berman · October 24, 2026 · Same story
- 13Kimi K3.1 Pro Open Weights Released by Moonshot AI
Matt Wolfe · November 6, 2026 · Same story
Follow this story
Open the full story →Moonshot's Kimi 3 to Rival Anthropic's Opus 4.8
16 developments
- Jul 16 · TechCrunch AI
Moonshot's Kimi 3 to Rival Anthropic's Opus 4.8 - Jul 17 · The Rundown AI
Moonshot AI's Kimi K3 Challenges Frontier Models↳ Kimi K3 features a 1M context window and will release weights by July 27. - Jul 18 · TechCrunch AI
Moonshot AI's Kimi K3 Sparks Global AI Debate - Jul 20 · AI News
Moonshot AI's Kimi K3 Model Bets on Memory (This article) - Jul 20 · The Verge AI
China Challenges US AI Dominance with New Models↳ Alibaba unveils Qwen3.8 alongside Kimi K3, joining the challenge to US AI dominance with open-source models. - Jul 20 · MIT Technology Review AI
China's AI Models Stir Controversy in Trump's Circle↳ Kimi sparks debate among Trump's AI advisors over open access versus national security intervention. - Jul 20 · TechCrunch AI
Debate Over Open-Weight AI Models Intensifies - Jul 21 · AI News
Chinese Open-Weight Models Spark U.S. Policy Debate - Jul 22 · Fireship
Moonshot releases Kimi K3 with 2.8 trillion parameters - Jul 22 · WIRED AI
Chinese AI Models Challenge Silicon Valley's Dominance - Jul 24 · WIRED AI
Silicon Valley Divided Over Chinese AI Models↳ Silicon Valley is divided on whether to regulate or allow free access to Chinese open-weight AI models due to security and IP concerns. - Jul 24 · WIRED AI
China-US AI Tensions Rise Over Model Distillation - Jul 27 · The Verge AI
China's Open AI Models Challenge US Dominance - Jul 28 · The Rundown AI
Moonshot releases largest open AI model K3 - Jul 28 · The Rundown AI
Moonshot releases largest open AI model K3 - Jul 29 · Together AI Blog
Together AI partners with Moonshot AI for Kimi models↳ Together AI partners with Moonshot to host Kimi K3, giving developers immediate access and fine-tuning capabilities.
More in Models & Labs
Models & Labsother
llama.cpp b12340 adds CUDA 13 and ROCm 10.0
This release quietly closes the hardware gap for local inference by adding default support for CUDA 13 and ROCm 10.0 across Linux and Windows. Users with newer NVIDIA GPUs or AMD accelerators no longer need to compile from source to get performance, while Apple Silicon builds now ship KleidiAI disabled by default, shifting the focus back to standard ARM optimizations. The inclusion of Snapdragon support for Linux and Android marks a significant step toward on-device AI beyond the x86/ARM desktop ecosystem. It is less about new features and more about ensuring llama.cpp remains the universal runtime for every available accelerator.
llama.cpp Releases
Models & Labscoding
llama.cpp b12342 adds few-row MMQ for speculative decoding
Speculative decoding just got a significant speed boost on NVIDIA GPUs. This release introduces a specialized kernel for 'few-row' matrix multiplications, targeting the small batch sizes typical of verification passes in speculative decoding workflows. By optimizing these specific operations for quantized formats like Q8_0 and IQ4 variants, llama.cpp reduces latency where it matters most during token generation. The change is CUDA-only for now but leverages hardware capabilities from Ampere onwards, making local inference faster without requiring new model architectures or complex configuration changes.
llama.cpp Releases
© TechCrunch AIModels & Labsmodels
Google deploys Gemini 4 Hydrogen for faster Search
Google is quietly shifting its search infrastructure by replacing older fast models with Gemini 4 Hydrogen, a tier designed specifically for speed and cost-efficiency rather than raw reasoning power. This model completes tasks 2.5 times faster than its predecessor while beating the more powerful Argon on specific intelligence benchmarks, proving that specialized efficiency can outperform brute force in high-volume workflows. By integrating this into Search's fan-out technique, Google now breaks down complex queries into subqueries with greater precision and speed, delivering richer visual responses without the latency of heavier models. This marks a strategic pivot toward agentic execution layers within consumer search, making AI Mode feel snappier for everyday multi-part questions.
TechCrunch AI