| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| release-info.json | 2026-10-18 | 2.0 kB | |
| vmlx-1.6.80.tar.gz | 2026-10-18 | 3.1 MB | |
| vmlx-1.6.80-py3-none-any.whl | 2026-10-18 | 3.4 MB | |
| vMLX-1.6.80-sequoia-arm64.dmg | 2026-10-18 | 577.9 MB | |
| vMLX-1.6.80-sequoia-arm64.dmg.blockmap | 2026-10-18 | 602.7 kB | |
| vMLX-1.6.80-tahoe-arm64.dmg | 2026-10-18 | 603.3 MB | |
| vMLX-1.6.80-tahoe-arm64.dmg.blockmap | 2026-10-18 | 630.0 kB | |
| README.md | 2026-10-18 | 2.4 kB | |
| vMLX 1.6.80 source code.tar.gz | 2026-10-18 | 11.8 MB | |
| vMLX 1.6.80 source code.zip | 2026-10-18 | 13.1 MB | |
| Totals: 10 Items | 1.2 GB | 0 | |
vMLX 1.6.80
Models and inference
- Add MiMo-V2.6 image, video, audio and tool support, bundled DFlash detection, and SSD-backed conversations. Restore ordinary decoding when the drafter is disabled.
- Add Nemotron-3.5 Bolt and Super-VL, including native vision support and compatible Super-VL MTP. LiLiCorr remains optional and off by default.
- Improve long-document prefix reuse, prefill cancellation, and background cache persistence. Preserve native model sampling and existing speculative decoding defaults.
Images
- Add Qwen Image 2.1 and Turbo generation/reference editing and compatible Ideogram 4 generation.
- Add native SeedVR2 resolution and softness controls, saved output dimensions, structured FIBO captions, and explicit Flux2 Klein editing.
- Preserve image adapter selection across requests, new sessions, and deep sleep; validate unsupported controls clearly.
App and API
- Keep earlier MiMo media in Responses follow-ups; return recoverable errors for rejected tool output without executing invented tools or automatically retrying generation.
- Fix remote-session switching, preserve string-valued GLM tool arguments, and keep large text pastes responsive.
- Explain why changing tools or thinking can invalidate a long chat's prefix cache. Show tool-loop prompt usage as a total across requests.
- Include JANG 2.5.50 Nemotron vision modules and qualified MFlux 0.23 / Pillow 12.3 dependencies.
Performance depends on the bundle, prompt, sampling, context and device state. Speculative decoding can benefit code or long-context workloads while remaining slower than ordinary decoding on some prose workloads; this release does not promise a universal speed floor.
Use the Tahoe build on macOS 26 or newer for native Metal kernels. The Sequoia build supports macOS 14.5 and newer.
Verified release artifacts
Built from 3830f403, with JANG 6dd1268e. Both DMGs are signed, notarized and stapled.
| Download | SHA-256 |
|---|---|
| vMLX-1.6.80-sequoia-arm64.dmg | 1f97895dc875b18655e33865233fe21a418eaad91bd5defdddaa4f8fcd0c56ea |
| vMLX-1.6.80-tahoe-arm64.dmg | 586dec74d543e5fe270cee3a07aa6ee575a7c9b0e03ce968f44ae2a50d50f878 |
Full provenance and Python distribution hashes are in release-info.json.