Kimi K3 Launches: Open‑Source 3‑Trillion‑Parameter Model Challenges Claude Fable 5

Open 3T‑Level Model

Kimi K3 is the world’s first open‑source large model with roughly 3 trillion parameters, supporting a 1 million token context window and native visual understanding. The model’s architecture is designed to meet the demands of ultra‑long sequences, deep network information flow, and high‑sparsity Mixture‑of‑Experts (MoE) routing.

Technical Composition

According to Kimi’s technical blog, the core components include:

Kimi Delta Attention (KDA) – addresses attention‑cost efficiency for long sequences.

Attention Residuals (AttnRes) – selectively retrieves representations from different depths instead of aggregating them uniformly.

Stable LatentMoE – expands total parameters to 2.8 trillion while activating only 16 of 896 experts per forward pass.

Quantile Balancing – routes experts based on the quantile of routing scores, reducing reliance on heuristic updates and sensitive balance hyper‑parameters.

Quantization‑Aware Training – uses MXFP4 weights and MXFP8 activations to accommodate low‑precision hardware.

Per‑Head Muon – extends the Muon optimizer to optimize each attention head independently, enabling more flexible learning at massive scale.

Sigmoid Tanh Unit (SiTU) and Gated MLA – enhance activation control and attention selection.

These innovations together allow K3’s overall scaling efficiency to be about 2.5 × that of its predecessor K2, providing a stable training and deployment pipeline for a model of this size.

Benchmark Performance

On the Arena.ai front‑end code arena, Kimi K3 ranks first, far ahead of Claude Fable 5. In the Artificial Analysis index, K3 achieves a score of 57, compared with 60 for the closed‑source Claude Fable 5 and 59 for GPT‑5.6 Sol, indicating a slight gap with the strongest proprietary models but a clear lead over other open models.

Although overall intelligence lags behind the top closed models, K3 consistently outperforms other evaluated models across the benchmark suite and demonstrates an excellent performance‑to‑cost ratio in practical use.

Knowledge‑Work Impact

K3 advances end‑to‑end knowledge work. Internal evaluations show continuous performance gains driven by patterns and challenges observed in real‑world agent workflows, reflecting a comprehensive improvement in the model’s ability to assist knowledge‑intensive tasks.

The model is publicly available, with the default reasoning intensity set to “Max”. Future updates will add “Low” and “High” modes.