Dark Side of the Moon open source Kimi K3: 2.8 trillion parameters are fully open, overseas developers call it "local version of Fable 5"

When the open source model community still regarded 300B as "big", Dark Side of the Moon directly raised the threshold to the trillion level. On July 27, Kimi K3 was officially open sourced - 2.8 trillion parameters, the world's first 3T-level parameter model open to the outside world, combining "trillion-level parameters" and "open source", two previously almost mutually exclusive words, together for the first time.

The gold content of three hard indicators

Kimi K3 refreshes the existing boundaries of the open source model in three dimensions. The first is the parameter scale: the size of 2.8 trillion means that it is not only the first open source trillion-level model in China, but also directly pushes up the capability ceiling of open source models by an entire order of magnitude. Second is the architecture: K3 adopts a mixed expert (MoE) architecture and uses the idea of ​​sparse activation to maintain usable reasoning efficiency under extremely large parameter scales. This is also the premise for trillions of parameters to truly "run". The third is capability coverage: the model has native visual capabilities and can handle multi-modal tasks of graphics and text end-to-end. The context window reaches 1 million tokens, which is enough to support ultra-long document understanding and complex long-term tasks.

Reaction from overseas communities

Feedback from open source comes quickly. After the model went online, it quickly topped the Hugging Face trend list. The overseas developer community gave a very informative evaluation - calling Kimi K3 "local version of Fable 5". The implication is that it provides an experience close to the cutting-edge closed-source model, but can be fully deployed locally and the weight is completely open. Elon Musk also liked and commented on social platforms, spreading this wave of heat from the developer circle to the wider public.

The background of this phenomenon is a subtle shift in the power balance between the open source and closed source camps. In the past, the open source model was usually regarded as a "sufficient but not top-notch" option; the emergence of Kimi K3 satisfies the two demands of "near cutting-edge capabilities + complete openness" for the first time on a trillion-level scale, and also caused cracks in the old narrative that "open source cannot catch up with closed source".

From a broader industry perspective, the domestic open source route represented by DeepSeek and Kimi is using "ultra-large scale + open weight" to take a completely different path from the closed source frontier - the former has the upper limit of the capacity, while the latter exchanges openness and localizable deployment for ecology and trust. It is worth noting that open source models with trillions of parameters place stringent requirements on inference infrastructure. The actual cost and ease of use of local deployment will determine how much real adoption this wave of "open source ceiling raising" can ultimately translate into.

Several directions worth tracking in the future:

1. Inference cost of Kimi K3 under real load: Whether the sparse activation advantage of MoE can offset the memory pressure of trillion-level parameters is the practical issue that the community is most concerned about.
2. Sustainability of the popularity of the HF list: The popularity of the trend list will eventually ebb. Whether K3 can precipitate a stable developer ecosystem is the long-term indicator.
3. The follow-up rhythm of domestic open source: Whether DeepSeek, Qwen, etc. will accelerate the layout of trillion-level open source, will determine whether the open source camp can raise the ceiling as a whole.
4. Follow-up to Musk’s statement: Does the xAI leader’s public praise indicate recognition or possible cooperation on a certain technical route?