Modular Says Max Beats vLLM 5.5x on Kimi K2.5 as Mojo 1.0 Stays Gated

Modular says its Max stack is 5.5 times faster on time to first token for Kimi K2.5 than vLLM and delivers 1.5 times more tokens per minute per GPU on Nvidia B200, according to benchmarks the company says were independently verified by Artificial Analysis. StartupHub.ai reported the claims, which Abdul Dakkak of Modular presented at AI Engineer.

Dakkak's Max and Mojo Claims at AI Engineer

Dakkak told AI Engineer that Modular is rebuilding the inference stack to hide hardware choice, describing the hardware market as a Cambrian explosion. At the bottom of that stack is Mojo, a Pythonic systems language he likened to C++, Rust or CUDA C++, targeting CPUs, GPUs and accelerators, where knowing Python covers about 90% of the syntax. He said Mojo competes with CUDA on performance and aims for portability without the usability tradeoffs of ROCm or OpenCL, while Triton hits about 70% of peak performance, and he pointed to more than 700,000 lines of kernel code the company says it has open sourced.

On top of Mojo sits Max, which Dakkak framed as a replacement for the PyTorch plus vLLM or SGLang or TensorRT-LLM patchwork for generative inference, with native KV cache handling, graph compilation and eager execution for debugging. He said Max supports thousands of architectures from Gemma 4 to MiniMax and GLM with an OpenAI-compatible endpoint, and can schedule prefill on one accelerator and decode on another across multiple GPUs and nodes. For image generation, he said Max is 7 times faster than Diffusers with compile on Flux 2, and 3 times faster than the next competitor on Flux 2 dev per Artificial Analysis, while on AMD MI355 he cited a 4 times throughput advantage over the equivalent setup, noting AMD parts are typically about 30% cheaper than Nvidia equivalents. Dakkak said the numbers are stated as raw speed without distillation or quantization tricks, and the company invites users to pip install modular to reproduce them.

Artificial Analysis launched AA-3D on October 12, 2026, a benchmark where AI models generate interactive three.js scenes from text briefs and human judges rank them via blind pairwise votes, following AA-Video-T2V v2.0.

Modular's Funding and Qualcomm Deal

According to akta.pro company records, Modular has raised $380 million across three funding rounds: a $30 million seed in June 2022, a $100 million Series A in August 2023 and a $250 million Series B in September 2025. The same records show a $1.6 billion valuation after the September 2025 round and a Qualcomm acquisition announced in June 2026 for approximately $3.92 billion in an all-stock deal expected to close in the second half of 2026.

Qualcomm, HTEC and ModCon Moves

Qualcomm announced at Snapdragon Summit 2026 that its AI software stack, built on Modular, supports six architectures from six companies, with Mojo, Max and Modular Cloud integrated into its Snapdragon portfolio, Nate News reported in September 2026. HTEC and Modular demonstrated Modular's AI software stack running on Google TPU at ModCon 2026, extending the MAX and Mojo ecosystem to the new hardware, FinancialContent Business Page reported in September 2026. Forbes reported in August 2026 that Qualcomm has fully open-sourced Modular's software stack following its $3 billion acquisition, committing to support competitors' silicon such as those from Nvidia and AMD.

What Remains Gated

Mojo 1.0 and full open source were described as imminent, not available, during the talk. The comparisons are presented as best versus best but without published scripts, batch sizes, precision, or price per token, and the only external validation named is Artificial Analysis on a single image model.