Moonshot has introduced Kimi K3, a 2.8 trillion parameter artificial intelligence model that reportedly completed the design of a functional inference chip during a 48 hour autonomous run.
The model handled the design, optimization, verification, timing analysis, and simulation process using open source electronic design automation tools. The resulting chip was created for a smaller Kimi K3 Nano model and reportedly achieved 8,721 tokens per second in simulation at 100MHz.
Moonshot describes Kimi K3 as an open frontier model built for advanced reasoning, coding, multimodal work, and long running agent tasks. It supports text, images, and video while offering a context window of more than one million tokens.
The chip design demonstration is one of several examples Moonshot is using to show how the model can complete complex technical projects with limited human involvement.
The chip uses a compact design built around INT4 inference
The reported chip measures approximately 3.981 square millimetres and contains around 1.46 million standard cells.
It includes 0.277MB of SRAM and an INT4 multiply and accumulate array with fused dequantization. These features are intended to support efficient inference for the smaller Nano model.
| Chip specification | Reported detail |
|---|---|
| Design time | 48 hours |
| Area | Approximately 3.981 square millimetres |
| Clock speed | 100MHz |
| Standard cells | Around 1.46 million |
| SRAM | 0.277MB |
| Numerical format | INT4 |
| Simulated throughput | 8,721 tokens per second |
| Design tools | Open source EDA software |
| Process library | Nangate 45nm |
The project used the Nangate 45nm library rather than a modern commercial manufacturing process. This means the result should be understood as a technical demonstration rather than a production ready processor intended for immediate manufacturing.
Simulation results also do not guarantee that physical silicon would achieve the same performance. Manufacturing variation, memory behaviour, power delivery, thermal limits, and packaging can all affect a real chip.
Even with those limitations, completing a full design flow in 48 hours would be a notable example of AI assisted hardware development.
Kimi K3 uses a mixture of experts architecture
Kimi K3 contains 2.8 trillion total parameters and uses a mixture of experts design. This architecture activates only part of the full model during each task, helping reduce the computing resources required for inference.
The model includes a context window of 1,048,576 tokens, allowing it to process very large documents, codebases, videos, and collections of related material during a single session.
Moonshot also introduced Kimi Delta Attention and Attention Residuals. The company claims these systems can improve decoding speed in very long contexts and increase training efficiency with limited additional computational cost.
| Kimi K3 feature | Detail |
|---|---|
| Total parameters | 2.8 trillion |
| Architecture | Mixture of experts |
| Context window | 1,048,576 tokens |
| Input types | Text, images, and video |
| Model variants | K3 Max and K3 Swarm Max |
| Reasoning mode | Always active with adjustable effort |
| Weight format | MXFP4 |
| Activation format | MXFP8 |
| Planned open weight release | Around July 27, 2026 |
Moonshot says the model will be available through its application, coding tools, compatible API, and third party services. Open weights are expected to follow later in July.
The company lists API pricing at $3 per million input tokens and $15 per million output tokens, with lower pricing for cached input.
The model also created a GPU compiler and edited its own launch video
The chip was not the only autonomous project shown during the Kimi K3 announcement.
Moonshot says the model created MiniTriton, a GPU programming system and compiler built from the ground up. The company claims it can match or exceed parts of NVIDIA’s Triton compiler in selected tests, although independent verification will be needed.
Kimi K3 also reportedly edited its own promotional video from 56 raw clips. The work included clip selection, continuity, music synchronization, audio processing, revision, and explanatory motion graphics.
These demonstrations are intended to show that the model can manage long, multi stage tasks rather than only generate short responses.
Moonshot is also promoting its ability to turn images and videos into interactive scenes. One demonstration generated a playable environment featuring a character riding a horse through a farm during rainfall.
This type of system could eventually help developers create prototypes, environments, simulations, and interactive media more quickly. However, early demonstrations may not reflect the quality, reliability, or control required for full commercial game development.
Performance claims still need independent testing
Moonshot reports strong results in coding, agent, and visual benchmarks, but the figures are self reported.
The company acknowledges that Kimi K3 remains behind some leading proprietary models in broader evaluations. Its main advantage may be the combination of large scale, open weights, long context support, multimodal input, and strong agent capabilities.
The chip design result is also based on simulation rather than independently tested manufactured hardware. It demonstrates a complete automated workflow, but it does not yet prove that the model can replace experienced chip engineering teams.
Kimi K3 still represents an ambitious step in open AI development. Its ability to complete hardware design, compiler development, video editing, and interactive scene generation suggests that future models may take on increasingly complex technical projects.
Moonshot plans to release the model’s open weights around July 27, 2026. That release will give researchers and developers a better opportunity to test its performance, study its architecture, and verify the company’s claims.
Discussion (0)
Be the first to comment.