vMLX 1.6.20

vMLX 1.6.20

vMLX 1.6.20 is a DeepSeek V4 Flash 0731-focused checkpoint for the Electron application and Python engine.

Highlights

• DeepSeek V4 Flash sessions now derive native reasoning, DSML tool support, sampling guidance, cache-tier capabilities, and pool-quantization behavior from the selected bundle and effective runtime.
• Server and Chat settings distinguish native DSV4 compiled decode and pooled-cache state from generic TurboQuant and unsupported whole-model cache modes. Activation-QAT can be controlled explicitly, with its effective state reported in the application.
• Native DSV4 cache handling now preserves short append checkpoints, lossless L2 writes, valid eviction ancestry, partial-tail replay, SSD-only operation, and RAM-to-SSD refault while avoiding repeated validation reads.
• Cache reporting separates current-process reads, writes, misses, and evictions from persistent namespace and managed-root storage. Session-scoped status requests prevent stale cache results from appearing after a session changes.
• DSV4 context admission, stop-token handling, native tool-prompt ownership, DSML failure behavior, and tool-result continuation were hardened across Electron and OpenAI-compatible API paths. Health output now distinguishes the served model from the currently loaded model.
• Explicit visible-final instructions after tool use now enter a no-more-tools continuation pass, and terminal abort cleanup is drained before the runtime returns to idle.
• Electron settings warn when a DSV4 top-p value differs from bundle guidance, preserve remote model tool contracts, and keep loaded-session navigation and cache activity displays readable and truthful.
• Release packaging now attests the native Python DSV4 encoder, nested affine defaults, mixed module quantization metadata, and the clean JANG runtime source. Separate Sequoia and Tahoe packaging/notarization contracts remain bound to one source revision.

Scope

This checkpoint is intentionally focused on DeepSeek V4 Flash runtime, cache, tool, settings, and packaging paths. It does not imply exhaustive certification of every model family, modality, or maximum-context configuration.

Thanks to GitHub contributor @hornsan1 for the decode-path improvement incorporated into this release.