PeytonT/SimulateAnything-Gemma-3-1B-v2

SimulateAnything Gemma 3 1B v2

Experimental derivative for typed scene and native-lab planning. The q4 artifact at ce7a00c3f0d4ffdce0c6ca6babdc2e410f484f47 was tested through actual WebGPU inference, strict compilation and native execution.

The adapter continues the v1 scene-authoring adapter with 2443 synthetic/replay examples and 1155 refinement examples. Native BF16 evaluation completed 34/52 development simulation requests and 9/11 additional paraphrases; no-mutation checks passed 9/10 and 3/3. Eighteen development requests have no executable chat adapter and remain incomplete. Refinement used development failures; these are not unbiased general intelligence scores. The additional paraphrases were frozen before refined outputs were inspected. Actual q4 browser completion was 31/52 development requests and 7/11 additional paraphrases, compared with v1 at 3/52 and 2/11 under the same expanded context. Browser no-action checks passed 9/10 and 3/3; the remaining development case emitted an invalid unsupported plan, which the application rejects. Three sorting actions regressed during q4 conversion. See qualification.json for exact cases, outcomes and limits.

Supported contracts cover bounded rigid-body scenes, visual forest/creek/lake layouts, nine existing physics models, sorting/search, and chip controls. Trees, scenery water and wind are visual effects, not a coupled hydrology/ecology solver. Model output can be wrong and must pass the application's typed validator. JSON validation does not establish user intent or physical accuracy.

The q4 ONNX graph uses block size 32. The adapter is also included. Base revision and training/export identities are in provenance.json. Use the application contract from the SimulateAnything repository; this is not a general-purpose code executor.

Gemma Terms of Use and prohibited-use restrictions apply. See LICENSE and NOTICE. Modified weights are identified in NOTICE.

The final application places current state before the latest request and accepts exact installed versioned environment IDs as typed aliases. Three changed panel inputs were rerun; other inputs were unchanged. V1 used the previous property order. Actual UI tests passed creek-to-lake replacement, automatic four-sphere creation, no mutation for questions/negation, and persisted previous-scene recovery. These are separate from the single-generation model scores.