Skip to releases

JSPARK3

Three Sparks. One model server.
The releases, recipes, and measured results.

Releases

  1. Latest release v2.0.2 · Oct 5, 2026GLM-5.3 Flash 92.5tok/sCode, one requestfresh prompt73.5tok/sProse, one requestprompt cached≈2,149tok/sReading a 32K-token promptfresh113.4tok/sFour at once, end to end (RigMark)fresh promptsbase weights + draft model · v2.0.1, measured October 3, 2026Recipe, results, and release notes
  2. Named release Experimental · Recipe v2.0.3Tempo DeepSeek-V4.1 Flash. Our DeepSeek experiment.Recipe, results, and limitations

Release history

  1. v2.0.2Patch release: image checkpoint reuse and GIF supportLatestOct 5
  2. v2.0.1GLM-5.3 Flash on a fork of TensorFold 0.3.6.2 (MIT)Oct 3
  3. v2.0.0Internal build, not published
  4. v1.8.4GLM-5.3 Flash · vLLMSep 28
  5. v1.8.0GLM-5.3 Flash · vLLMSep 27
  6. v1.2 to v1.7Internal builds, not publishedSep
  7. TempoDeepSeek-V4.1 Flash · recipe v2.0.0 to v2.0.3Sep 13
  8. v1.1 CadenceGLM-5.3 FlashSep 7
  9. v1.0GLM-5.3 FlashSep 2

Numbered JSPARK3 releases are the main line, and it runs GLM-5.3 Flash. Tempo is a named release: a DeepSeek experiment with its own recipe versions, numbered apart from the main line. Links to earlier GLM releases lead to the v1.8.4 page, kept as published.