Mon Shot makes Kemi 3.1 Pro weights available: 2.8 trillion parameters for agents’ self-run recalculations

Mon Shot AI has made the weights of its new model “Kemi K 3.1 Pro” available, offering developers a multi-modal agent model with open weights totaling 2.8 trillion parameters. The model uses a mixture-of-experts architecture, activating only 104 billion parameters per token, based on a massive context window that fits one million code tokens, followed by large-scale reinforcement learning across hundreds of specialized domains.

A fundamental change has been made to the attention architecture to enable more efficient processing of long contexts.The company replaced the constrained multi-head attention layers with sparse attention layers it calls “Sparse DeepSec Attention”, using 24 layers, together with 69 layers from the KDA architecture, relying on 896 experts from which 16 experts are selected per token plus two shared experts. This architecture was merged with a visual encoder named “MonFitV2” containing 401 million parameters, and trained with quantization that provides weights at MXFP 4 precision and activations at MXFP 8 precision to ease inference on engines such as FLLM, SG-Lang and TokenSpeed.

Evaluation results show a marked advantage on programming and autonomous-agent tasks, with the model scoring 65.7 points on the Terminal Bench 4.0 command-line test and 75.5 points on the Deep SW-E software-engineering benchmark, outperforming leading closed-source models on the same tests. In simulated browsing and workplace-automation environments it recorded 92.1 points on the browsing evaluation and 73.3 points on business automation, with a noticeable drop in code-hallucination rates to 8.3 points on the hardest tool-augmented tests, excelling in function-call accuracy and extended reasoning chains that require full preservation of inference texts and tool requests across conversation turns.

Possessing weights of this size redraws the boundaries of digital sovereignty and operating economics in the region.I observe in this proposal a tangible path for financial institutions and government data-centers in Saudi Arabia, the UAE and Qatar to move beyond exclusive reliance on foreign cloud APIs and the rising dollar cost per million tokens. Local tech teams can now deploy programming and analytical capabilities that match major closed models on their own servers, meeting data-governance and banking- and health-data confidentiality requirements without sending sensitive inputs to external servers.

At the level of development teams and emerging tech firms in Egypt and Jordan, the shift to open-weight trillion-parameter models changes cost structures and skill requirements. Instead of paying bills for proprietary API consumption, the technical effort focuses on engineering low-cost local inference pipelines that rely on four-bit quantization and distributed computing, enabling the building of advanced agent systems that can read codebases of hundreds of thousands of lines without draining early-stage operating budgets.

But moving to this path requires a careful read of infrastructure calculations, as the model licensed under “Kemi K 3.1 Pro” needs processing matrices capable of handling trillions of parameters even with weight quantization. Verify your infrastructure’s readiness and actual memory cost before deciding on self-hosting, since real benefit remains contingent on the ability to manage the software engines flexibly and ensure stability of complex inference pipelines.