S-Space is a research codebase released by the MirroS Team that studies whether multimodal models maintain an internal spatial workspace and, if so, how to read and manipulate it. The authors report identifying a three-dimensional linear subspace inside intermediate activations, named S-Space, that encodes continuous object positions along horizontal, vertical and distance axes; per the accompanying blog and technical report, this representation persists across prompts and multi-view contexts, integrates information about objects outside the visible frame, and can be causally edited to shift a model's spatial judgments. The release ships pretrained S-Space axes plus experiment launchers for five backbone models -- Molmo2-ER, MolmoAct2-Pretrain, MolmoAct2, Qwen3.5-4B and Qwen3.6-27B -- with code organized under sspace/ for core logic and configs/ for experiment settings, installed via a uv-based script that also downloads the base models and InstructPart auxiliary models. Included experiments span constructing S-Space axes on COCO; comparing native model answers with S-Space coordinate readout on CV-Bench, EmbSpatial and SpatialTunnel; testing viewpoint-following on the HStar multi-view benchmark; running causal interventions that edit S-Space coordinates on EmbSpatial while checking non-spatial judgments stay intact; comparing part-whole spatial judgments across the three vision-action backbones on InstructPart; and contrasting direct and chain-of-thought answers with S-Space readout on SpinBench, with and without an added premise. On the SpinBench perspective-taking task, the accompanying blog reports that explicit computation over S-Space reaches 94.52% accuracy for Molmo2-ER, compared with 60.96% for the same model's direct answer generation, illustrating the project's central claim that models hold spatial information they do not always use reliably on their own. The repository was published alongside a report PDF and blog post on September 7, 2026, and is licensed under Apache-2.0; the benchmarks it depends on -- SpatialTunnel, EmbSpatial-Bench, CV-Bench, HStar and SpinBench -- and the base models it evaluates are credited to their original authors rather than to the MirroS Team.