vMLX 1.6.49

vMLX 1.6.49

Changes

• Publish vMLX 1.6.48 updater manifest (507b3794)
• Harden release publication and recovery automation (644666bc)
• Update release Actions to current immutable pins (a8662de3)
• Bind native MTP benchmarks to exact source (41d735ee)
• Fuse GLM MTP verify hyper-connections (d916a836)
• Gate GLM verifier slab fusions behind opt-in (c250fbf6)
• Align GLM native MTP head cache experimentally (a905c45d)
• Prime GLM native MTP head from prompt history (72a2375b)
• Make GLM prompt priming request-owned (7c2c3a25)
• Report GLM prompt priming capture failures (b2109637)
• Capture GLM prompt history at prefill owner (1b0f94c1)
• Capture GLM final prompt token before MTP seam (ea43f770)
• Fallback text MTP to AR on measured cost (77ad9671)
• Default GLM native MTP to autoregressive (0495d7a4)
• Add opt-in compact GLM MLA cache path (c1426652)
• Report active GLM cache schema from runtime (5c5c5482)
• Default GLM to compact absorbed MLA cache (a270350b)
• Advertise compact GLM cache as the family default (67d085fa)
• Amortize compact GLM MLA cache appends (64230805)
• Preserve stochastic native MTP sampling semantics (67c059f0)
• Preserve sampling in Auto native MTP (b5698d42)
• Align UI proof with Auto MTP sampling (3aacf02e)
• Grade effective UI sampling persistence (097f60ac)
• Support BF16 affine metadata in routed decode fusion (651605cf)
• Run routed decode fusion on BF16 activations (a4dba797)
• Separate adaptive MTP profiles for tool workloads (7c9d6b66)
• Support subfolder-backed image model artifacts (092719f1)
• Add native GLM-5.3 multimodal runtime (b9e893a3)
• Route embedded GLM VLM quantization metadata (f6e25093)
• Preserve GLM multimodal processor dispatch (cd08aa56)
• Normalize VLM hidden states for native MTP (d227fe73)
• Enable GLM multimodal sessions in the panel (37429726)
• Attribute MTP verifier acceleration per request (fb87e022)
• Retain affine MoE fusion on compatible layers (724ad7f4)
• Bind live proof to request reasoning and MTP records (b1878372)
• Grade raw reasoning against request mode (1a49ca12)
• Add sampled controls to native MTP benchmark (47802892)
• Prototype a proposal-only Qwen MTP head (13176274)
• Bound Electron proof target polling (d688b60e)
• Keep Qwen4 fixed-D3 drafting profitable (6470cd87)
• Fix Codex coding-tool configuration (0fea9253)
• Preflight local model bundle integrity before load (aeab349e)
• Keep bundle check JSON output machine readable (63498c39)
• Report bundle index repairs only when performed (856855e6)
• Gate every inference load on bundle integrity (dff701da)
• Prepare vMLX 1.6.49 (a972a297)
• Keep release contracts compatible with bundle preflight (493aa25e)
• Honor declared DSV4 release evidence deferral (ac9779fd)
• Hold tool terminals until cache persistence completes (aa744d53)
• Keep native Qwen tool prompts stable across continuations (9dab9f73)
• Trace cache block identity at store and restore (00fd5241)
• Add stable auto tool policy to agentic cache proof (bddc7812)
• Model stable Responses instructions in agentic proof (9fc66e96)

Runtime provenance

• vMLX source: 9fc66e966e18e501aab30683defab830a2864eb2
• JANG runtime: f3c79081c3d99e8a26fa28444934f7567acb00c6 (jang 2.5.47)
• Tahoe/macOS 26 is the default build; Sequoia is the compatibility build.
• Both DMGs are Developer ID signed, independently notarized, stapled, and Gatekeeper checked by the candidate workflow.

Downloads

• Tahoe: vMLX-1.6.49-tahoe-arm64.dmg
  • SHA-256: 7e0cf6b5e4b102b4e4fc1015a47cd4ab0024d13e5f4891efccc62dbfee9af1a5
  • Size: 538129484 bytes
• Sequoia: vMLX-1.6.49-sequoia-arm64.dmg
  • SHA-256: 1ac3d724b79c0a20c808340035f6f441ffe3f2b7138432ad7e51214e3acdce10
  • Size: 516049109 bytes

Python distributions

• Wheel: vmlx-1.6.49-py3-none-any.whl — SHA-256 b594c09f2c3dbe4ae0a2d607fcef6b90a5bd43fa471d93430b47ce514db6d0c0
• Source: vmlx-1.6.49.tar.gz — SHA-256 eecef6464e78950325ffdb301c4e373806b60b73c2bc2c050795656a2b9d7b73