LM Studio 0.4.26: runtime 2.50.0 (llama.cpp b11337), CUDA 13 on Windows ARM

LM Studio 0.4.26 is dated October 13, 2026 on LM Studio's changelog. It shipped in five builds. Build 1 lists only bug fixes and security hardening, with no detail. Build 2 updates the llama.cpp runtime to 2.50.0, which LM Studio says is upstream `b11337`. Build 3 adds CUDA 13 support on Windows ARM, UI polish, and coexistence with Bionic: with Bionic 1.1.8 or newer, both apps can run together, and Bionic uses LM Studio for local model inference while LM Studio is running. Build 4 adds a `/v1/systemone` endpoint, described as Jev (TypeSafe AI) compatible, and a `/v1/decisions` endpoint, described as OpenAI compatible. Build 5 fixes CUDA not being detected on Windows on ARM and fixes `lms` compatibility with Bionic. The changelog does not say how `b11337` relates to newer llama.cpp releases. On llama.cpp's GitHub, `b11337` was published on October 2, 2026. That is before the v0.6.0 tag (`b11429`, October 5) and v0.7.0 (`b11669`, October 15). `b11337` already has part of the v0.6.0 work, such as GLM-5.3-Flash support and MTP for Qwen4Exp. Other v0.6.0 items were merged upstream after `b11337`, including the server's `/v1/systemone` API, Clef and Nimble decision-model support, and new Metal flash attention and mat-mul kernels. All of v0.7.0 also came later. The changelog does not say whether LM Studio adds anything on top of `b11337`, or which models its new `/v1/systemone` endpoint supports. LM Studio gives no performance figures for this release, and we have not benchmarked it.