Moonshot has unveiled its latest Kimi K3 Open model, which packs a stunning 2.8 trillion parameters & was able to build a chip that crunches 8700 tokens/s in inference.
China's Kimi K3 AI Model Autonomously Designed A Fully Functional Chip That Delivers Over 8700 Tokens/s Through Inference Workloads
Today, China's leading AI firm, Moonshot, unveiled its brand-new Kimi K3 model, which is the world's first 3T-class "Open" AI model, spanning 2.8 trillion parameters, designed specifically for frontier-scale intelligence.
The technical highlights of Kimi K3 include:
- Parameters: 2.8 trillion total (Mixture-of-Experts / MoE architecture)
- Context window: 1,048,576 tokens (1M)
- Multimodal: Native support for text + image + video
- Architecture highlights: Innovations including Kimi Delta Attention (KDA) + Attention Residuals (claimed up to 6.3× faster decoding at million-token contexts)
- Weights format: MXFP4 weights / MXFP8 activations (quantization-aware from training)
- Variants at launch: K3 Max (general/chat/agent) and K3 Swarm Max (large-scale parallel processing)
- Reasoning: Always-on (with configurable effort via API at launch)
- Open weights: Promised release by ~July 27, 2026 (first open ~3T-class model)
- Pricing (API): $3 / $15 per million tokens (input/output); cached input is significantly cheaper (~$0.30)
- Availability: Kimi app, Kimi Code, Moonshot OpenAI-compatible API, OpenRouter
- Benchmarks (self-reported by Moonshot): Strong in agentic/coding tasks (e.g., Terminal-Bench 2.1: 88.3, just behind GPT-5.6 Sol). Competitive overall but trails top closed models in broad evaluations.
The AI model is said to trail behind flagship solutions such as OpenAI's GPT 5.6 and Anthropic's Claude Fable 5, but the model is available today and delivers some interesting highlights, such as a chip design that it was able to complete autonomously within 48 hours. The main highlights of the AI model include:
- Autonomous chip design — The model designed a real, functional chip (using its own architecture) in a single 48-hour autonomous run with zero human intervention in the design. It handled design, optimization, verification, timing, and simulation. The resulting chip (4 mm² area, Nangate 45nm library, open-source EDA tools) achieved > 8,700 tokens/second decoding throughput in simulation.
- Built a GPU compiler from scratch — created MiniTriton (a custom GPU programming system/compiler) entirely from scratch. It reportedly beats or matches parts of NVIDIA’s official Triton compiler on benchmarks.
- Professionally edited its own launch video — It took 56 raw video clips and edited a polished teaser, handling clip selection, action continuity, frame-level beat synchronization to music, audio processing, and multiple revision rounds. It also reportedly created 3Blue1Brown-style motion graphics explaining its own architecture.
In the various public benchmarks, Kimi K3 offers competitive and faster performance than the competition thanks to broad improvements in Kimi K3's agentic knowledge work capabilities, enabling more capable and reliable performance in real-world use cases.
8700+ Tokens/s Chip Made In Just 48 Hours
One of the highlights of Moonshot's Kimi K3 is that it was able to build a chip using its own model for its own Nano model. The chip is based on Kimi's own architecture, and was completed within a 48-hour autonomous run, which included the tape-out of the silicon "K3 built", optimization, and verification using open-source EDA tools (Nangate 45nm library).
The chip itself measures 4mm2 and clocks in at around 100 MHz. It packs 1.46 million standard cells, 0.277MB of SRAM, and an INT4 MAC array which features fused dequantization. In initial tests, the chip was able to achieve over 8700 Tokens/s of AI inference performance (8721 Tokens/s to be precise).
The Future of Game Dev & Digital Content Creation
Kimi K3 also promised to deliver strong and exceptional results for 3D reasoning, coding, & vision capabilities. It can be used to transform images and videos into fully playable interactive experiences, paving the way for future games where entire experiences can be built upon using a reference image or video.
The reference demo shows a character riding a horse through a small farm with trees in the back of the environment and a small lake. Rain is present in the scene, and the model does a great job in generating the scene, though this is just an early example. Moonshine shares several examples where images and videos can be turned into playable and interactable experiences.
Some of the examples of generated games/experiences can be seen below:
Kimi K3 represents a major leap in open-source AI, delivering frontier-level intelligence in long-horizon coding, multimodal reasoning, and complex knowledge work through its innovative 2.8T-parameter architecture.
As the world's first open 3T-class model, it narrows the gap with leading proprietary systems while excelling in agentic tasks—from GPU kernel optimization and full compiler development to interactive research visualizations and creative video editing. With native vision, a 1-million-token context, and broad availability today (full weights releasing July 27, 2026), Kimi K3 empowers developers, researchers, and creators with unprecedented open access to advanced AI capabilities.
About the author: A Software Engineer by training and a PC enthusiast by passion, Hassan Mujtaba serves as Wccftech's Senior Editor for hardware section. With years of experience in the industry, he specializes in deep-dive technical analysis of next-generation CPU and GPU architectures, motherboards, and cooling solutions. His work involves not only breaking news on upcoming technologies but also extensive hands-on reviews and benchmarking.
Follow Wccftech on Google to get more of our news coverage in your feeds.
