vMLX 1.6.19

vMLX 1.6.19

vMLX 1.6.19 is a focused Electron and Python-engine checkpoint for model-owned
generation settings, persistent cache correctness, native MTP diagnostics,
prompt accounting, and macOS packaging.

Highlights

Model-owned settings and API behavior

• New and restored Electron chats now hydrate temperature, Top P, Top K,
  repetition penalty, and output limits from the selected model bundle without
  presenting missing values as saved zero-valued overrides.
• Explicit per-request API parameters remain authoritative. Ollama requests now
  preserve an explicitly supplied top_k, including 0, instead of silently
  dropping it during request translation.
• Laguna sampling remains metadata-owned. vMLX does not inject a hidden
  Laguna-specific Top K value; the app warns when a Laguna XS 2.1 bundle does
  not declare its expected Top K 20 setting.
• Automatic prompt limits are now derived from the configuration of the model
  that is actually loaded, bounded by the model's declared context ceiling and
  current memory budget. Text-only prompt accounting also follows the
  tokenizer's beginning-of-sequence behavior.

Cache reliability and performance

• Persistent cache payload encoding is moved off the inference-completion path
  after tensors are safely detached from MLX, reducing synchronous CPU work at
  the end of a request.
• SSD cache publication and eviction preserve the causal parent chain required
  to restore a reusable partial prefix. Incomplete chains fail closed instead
  of exposing an unrestorable suffix as a cache hit.
• Memory-pressure trimming keeps the largest safe block-aligned prefix and
  reports reuse against the prefix that was actually consumed.
• Hybrid TurboQuant batch splits retain MLX dtype metadata, avoiding retry
  loops and stalled streams caused by malformed copied cache state.
• MiniMax M2-family automatic cache storage uses the correctness-first q8
  policy. Bundle-owned calibrated TurboQuant settings and explicit user
  selections remain authoritative.
• Cache namespaces for compatible looped-transformer bundles include the
  effective repeated-layer layout, preventing reuse between incompatible loop
  counts.
• MiniMax M3 native sparse-cache blocks are scoped to the prompt-prefill shape
  and its N-1 prompt-boundary contract. This prevents a token-identical prefix
  produced under a different prefill shape from restoring numerically different
  Lightning Indexer state.

Model and runtime hardening

• Qwen native-MTP loading now keeps already converted backbone normalization
  tensors intact while converting MTP-owned tensors independently.
• Native-MTP health and profiling expose bounded cache-lifecycle, acceptance,
  and phase-timing telemetry without changing generation policy.
• Compatible Nanbeige looped-transformer JANG bundles now validate their loop
  count, effective cache-slot count, and runtime metadata before generation.
  Inconsistent bundles are rejected rather than running with a silently short
  cache.
• DeepSeek-V4 Flash preserves ratio-zero SWA rings and lossless q8 CSA/HCA pool
  segments through prompt snapshots and SSD reconstruction, and fails closed
  when its native composite-cache state is incomplete or misaligned.
• Bounded Electron tool workflows better enforce requested exactly-once tool
  calls while leaving other explicitly requested tools available until their
  results have been returned.

Electron experience

• Chat Settings now show explicit loading, fallback, partial-metadata, and
  unavailable states instead of inventing sampler values while bundle defaults
  are being resolved.
• KaTeX rendering better distinguishes literal currency from adjacent inline
  math and prevents an unclosed dollar delimiter from consuming later prose.
• New settings and compatibility messages are localized in the shipped
  English, Spanish, Japanese, Korean, and Chinese interfaces.

Distribution

• The Electron app and Python package are versioned together as 1.6.19.
• The bundled runtime requires JANG 2.5.37.
• Separate Apple-silicon downloads are provided for macOS Tahoe and for
  Sequoia-compatible systems. Choose the build that matches the target macOS
  version.

Compatibility notes

• Model-bundle metadata remains the source of truth for default sampling,
  reasoning, parsers, modalities, and calibrated cache policy. vMLX does not
  hide an incomplete bundle by coercing prompts or silently clamping sampling.
• Laguna XS 2.1 publishers should independently validate corrected artifacts
  for long-reasoning repetition and include the intended generation defaults
  in the bundle metadata.
• Looped-transformer support requires the bundle and loaded runtime to agree on
  the repeated-layer cache layout.
• This is a checkpoint release, not an exhaustive certification of every model
  family, modality, maximum-context workload, cache-pressure scenario, or
  performance profile.