Chinese startup Moonshot AI has officially announced Kimi K3, the largest open-source language model to date. With a claimed 2.8 trillion parameters, native vision, and a context window of 1 million tokens, K3 is vying for the title of the most powerful open-source solution. According to internal team evaluations, in the overall ranking, the newcomer only trails behind closed systems — Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol.
Architectural Innovations and Efficiency
At the core of Kimi K3 lies the new Kimi Delta Attention (KDA) architecture and the Attention Residuals (AttnRes) mechanism. KDA optimizes the processing of long data sequences, while AttnRes allows the model to selectively extract critically important information from different network layers, rather than processing everything uniformly. This ensures a deeper understanding of context and preserves meaning even when working with complex and convoluted texts.
A key element of scaling is the sparse Stable LatentMoE architecture. Out of 896 available experts, only 16 are activated for each inference. This selectivity, combined with new data and optimized training methods, has improved the model's efficiency by 2.5 times compared to the previous K2 version. Notably, for nine of the last twelve months, models from the Kimi family have held the record for size among all open-source counterparts.
Comparison with Competitors: Where K3 Excels and Where It Falls Short
On several benchmarks, K3 demonstrates superiority, even surpassing renowned competitors. In the SWE Marathon test for solving long engineering tasks, the model scored 42 points compared to 35 for Fable 5 and 39 for Sol. In the BrowseComp web research task, the result was 91.2 versus 88 and 90.4, respectively. K3 also leads in the Program Bench test (77.8 versus 76.8 and 77.6) and achieves a near-maximum score in Terminal Bench 2.1 (88.3 points), trailing Sol by only 0.5 points.
However, on the FrontierSWE and HLE-Full (a complex multidisciplinary exam) tests, K3 lags behind Fable 5. It is important to note that testing methodologies varied: some models were run via Claude Code, others via Codex or Kimi's own KimiCode toolkit. This makes direct comparison not always accurate, but the overall picture is clear: K3 is an extremely powerful tool for agentic tasks.
Coding, Engineering, and Autonomy
K3 can conduct lengthy engineering sessions with minimal human intervention. In a test optimizing GPU kernels, the model operated autonomously for up to 24 hours, showing results on par with Fable 5 and significantly outperforming Opus 4.8, GPT-5.6 Sol, and GPT-5.5. Special mention goes to K3's ability to write MiniTriton from scratch — a compact compiler for GPU kernels with its own IR layer on top of MLIR.
Equally impressive is the chip design demonstration: K3 independently designed a neural network chip on its own architecture. Over 48 hours of autonomous work, using open-source tools, the model created a chip design with an area of 4 mm², a frequency of 100 MHz, and a performance of over 8700 tokens per second in simulation.
Knowledge Handling and Multimodality
K3 demonstrates a deep understanding of scientific data. In one case, the model reproduced complex formulas from computational astrophysics in two hours, a task that would take an experienced researcher several weeks. In another example, K3 prepared an interactive report on the 42-year history of the ASIC industry, processing over 11,000 pages of documents.
Thanks to native video processing, the model can edit clips, synchronize them with music, and handle audio. This opens new horizons for automating creative processes.
Limitations and Availability
The Moonshot AI team honestly acknowledges that K3 is sensitive to losing its reasoning history when switching agent environments and may exhibit excessive initiative in ambiguous situations. In terms of usability, it currently lags behind the flagships Fable 5 and GPT-5.6 Sol.
The model is already available on the Kimi website, in Kimi Work and Code services, and via API. The maximum reasoning mode is activated by default. Full weights and a technical report will be published on July 27.
Expert Opinion: The release of Kimi K3 is not just another record in parameter count. It is a demonstration that open-source models can compete with closed-source giants at the highest level, especially in agentic and engineering tasks. However, the gap in usability and sensitivity to context changes remind us that "raw" power is only part of the equation. The race for AI leadership is just beginning, and we are seeing the open-source sector accelerate its pace.