vMLX 1.6.27

vMLX 1.6.27 for macOS

A new flagship model family with vision and video, a prefix cache that finally uses the capacity it advertises, and a set of API-fidelity fixes — on top of the decoding-speed rollup from 1.6.26.

Muse Glimmer 30B — vision, video, and real reasoning controls

Muse Glimmer 30B runs end to end: its text tower, its windowed vision tower, its recipient-routed reasoning rail, and its ATEM tool dialect. All three bundles (JANG_2D / 4M / 6M) are supported.

• Vision: single- and multi-image prompts are described correctly — a red circle and a blue square in the same turn are named as such.
• Video: clips are read in temporal order — a clip that fades from red to blue is answered "red blue", not the reverse.
• Reasoning depth is now adjustable from the app. Chat Settings previously offered only the Auto/On thinking toggle for this family, so the four depths the model actually has were unreachable. It now shows Low / Medium / High / Extra High, and changing the setting measurably changes how much the model thinks.

The prefix cache now uses the capacity it advertises

• A long prompt no longer falls off the end of the cache. Sessions were created with an index that addressed only about 64,000 tokens, so a longer conversation could report no reuse at all on an exact repeat — and run slower than starting cold. The index now covers 262,144 tokens, and existing sessions are migrated automatically.
• Vision models no longer run a slower cache path in the app than on the command line. The app could disable paged RAM for a model that the engine expected to run paged, silently costing reuse. The two now agree, and the migration meant to correct existing sessions — which never actually ran, yet marked those sessions as already migrated — is fixed.

Changing reasoning depth mid-conversation keeps its own cache lineage: each depth reuses its own chain, and switching back to a previous depth picks that chain up again rather than starting over.

API fidelity

• Anthropic thinking.budget_tokens now caps the reasoning chain. It previously set only a template hint that no runtime consumed, so the budget did nothing.
• Ollama think string levels now work. Clients sending think: "low" | "medium" | "high" were silently ignored — thinking never even engaged. A level now enables thinking and selects the effort.
• No control markup in Muse answers. A recipient marker that reached the visible answer on a malformed or truncated stream is now scrubbed from what you see, while reasoning text and tool-call bodies are left byte-for-byte intact.

Carried forward from 1.6.26

Health polling no longer stalls generation on any model family, and DeepSeek V4 Flash holds smooth long-context decode — measured this cycle at 35 tok/s short-context and 32.6 tok/s with 431 tok/s prefill at 3k context, with cross-turn reuse verified exact.

Separate signed and Apple-notarized Apple-silicon downloads are available for macOS Tahoe and Sequoia-compatible systems. Tahoe is the default download.