Skip to content

vMLX 1.6.7

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 13 Jul 05:23
· 1911 commits to main since this release
v1.6.7
82ca028

vMLX 1.6.7

Highlights

  • Paged KV cache now defaults ON for autodetected text families (dense + hybrid-SSM/linear-attention: Qwen3.5/3.6, Zaya, Nemotron-H, LFM2.5, Laguna, Hy3, Step-3.7) when continuous batching + prefix cache are active. Multimodal/VL bundles, MiniMax-M3, openPangu-v2, and Gemma 4 keep their own cache paths; DeepSeek-V4-Flash keeps its composite opt-in. Explicit --use-paged-cache / --no-paged-cache always win. The Electron panel is fully reconciled — the visible cache toggle, the launch flag, and the engine's effective policy match for every family.
  • Reasoning defaults ON for every reasoning-capable family (family guards — Mistral none/high, MiniMax custom-off, Laguna template-default-OFF — remain authoritative). max_thinking_tokens honored + gated on streaming and non-streaming.

Fixed

  • Never-empty answer after runaway reasoning: a reasoner that consumes the whole token budget inside its hidden rail still streams a visible answer instead of empty content, across chat and /v1/responses, for qwen3.5/3.6, DeepSeek-V4, Step-3.7 and MiniMax. Answer pass runs a fresh context for deepseek_v4/step3p7; no reasoning/<think> re-entry leaks into the visible answer.
  • MLLM/VL paged cache now honors the same RAM byte ceiling as the text path (--cache-memory-mb/--cache-memory-percent); previously bounded only by block count. Verified live: Step-3.7 held flat at +0.27 GB RSS over an 18-turn multiturn.
  • Memory-cache lock inversion (community #233 deadlock), a stop_token_ids crash, and a lockless cache-clear race.
  • Paged cache RAM byte-ceiling eviction (Wave-18) + disk-L2 longest-prefix off-by-one.
  • Loaded MLLM /v1/completions (+stream, +Ollama raw) route through the chat rail (fixes Gemma-4/VL constant-token degeneration).
  • gpt_oss streaming no longer leaks harmony analysis into visible content.
  • Laguna uses the glm47 tool parser; ZAYA/ZAYA1-VL pin zaya_xml.
  • Truthful /v1/capabilities cache/TQ descriptors; panel launch RAM estimate no longer over-counts lazy-mmap JANG bundles.

Notes

  • Step-3.7-Flash video verified engine-sound (clean reasoning/answer split under adequate output budget; it is a very long reasoner).

Downloads: notarized macOS DMGs at https://github.com/jjang-ai/mlxstudio/releases/tag/v1.6.7 — Sequoia (Sonoma 14.5+/Sequoia/Tahoe) and Tahoe-native.

Assets 2
Loading