Repository navigation
Release Notes
Outcome: Find what changed, which build the instructions describe and what has actually been tested. LexiSuite is in active development, not a generally available, production-qualified release. Versioned Panel changelog entries and USB image revisions are different release streams; an image name should not be mistaken for a Panel package version.
- File:
LexiSuite-14d173b-secureboot-dev.img.xz(351,058,120 bytes), with its.img.xz.sha256(66d1c50878a000cf6949b4dc6f2ea4eeb590fcd07e6e6281f651e2950abf02a0) and the decompressed.img.sha256(1bb47e45fc14dfc4eea1c7a6b64f65650e8a746bd89c1fa92cb71383225f0c0b). Linux entry 7.4-rc2, as inafaf544. - The newest image you can download: from the download page.
- Engine: Gemma 4 prompts are read about twice as fast. Layers that reuse another layer's keys and values run only the rows that produce output, attention loads each key/value row once for the query heads that share it, and prompt passes are 512 rows. Output is identical bit for bit. Gemma 4 E4B on the RX 7900 XTX: 556 tokens 1,812 to 3,782 tok/s, 3,595 tokens 2,300 to 5,688 tok/s (llama.cpp 1.0.0: 3,370 and 5,923).
- Engine: a wrong result on Gemma 4 prompts longer than the sliding window is fixed. The local key/value ring now holds the window plus one prompt pass; a device test passes with the fix and fails without it.
- Engine: Gemma 4 31B is supported. Default settings on the RX 7900 XTX: prompts 561 / 684 tok/s against llama.cpp 1.0.0's 469 / 594, generation 35.4 / 34.7 / 33.7 against 37.2 / 36.2 / 34.8 tok/s.
- Engine: a 4-bit K/V cache for Gemma 4 and contexts up to 262,144 tokens. The memory plan prints the largest context that fits the card (about 205,000 tokens for the 31B on 24 GB). At a 210,176-token context the 31B reads a 3,595-token prompt at 817 tok/s and a 32,966-token prompt at 548 tok/s, against 661 and 447 for llama.cpp 1.0.0. Known fault: on the 31B a prompt beyond about 130,000 tokens stops with an out-of-memory refusal on its last pass; the correction is in test for the next version.
-
Engine: Gemma 4 generation stores keys and values in 16 bits. Identical output, half the cache, about 1.5 GB less video memory on the 31B. E4B generation 135 / 129 / 125 to 139 / 134 / 131 tok/s.
LEXI_GEMMA4_NORM_CHAIN=1(off by default) takes the 31B to 38.9 / 37.9 / 36.8 tok/s. - Engine: the Gemma tokenizer is 14 times faster on a long prompt, same tokens.
- Rust 1.100.0 is the pinned toolchain; the gates and the image build ran on it.
- LexiOS, LexiKernel and the Panel are as in
afaf544. RTX 2060 numbers are unchanged. - Tests on its source: engine 4,164 passed and 1 failed (a timing race inside one VM-memory test; it passed on three reruns and is corrected in the next version), GPU server 570, the CUDA kernel suite on the card 185, LexiOS, LexiKernel and Panel as before. The image passed the nine boots of the VM qualification.
- File:
LexiSuite-afaf544-secureboot-dev.img.xz(350,967,784 bytes), with its.img.xz.sha256(ee790fcf10dcf0c26254c9f47229bb1f17ddb4c49c51ce839af63053605d5869) and the decompressed.img.sha256(24aae3d351bb63aed581faf385bfc1c399f21745958924804b8f768209e26676). Linux entry 7.4-rc2, as in5db82d6. - Superseded by
14d173babove. -
Engine: long prompts on the 4-bit cache are read up to 33% faster. A new prompt attention kernel packs the query heads that share a key/value head so no GPU subgroup idles and stages each key/value tile 24 times instead of 32. Same arithmetic, output identical bit for bit. On a 27B model at a 262,144-token context: 33,000 tokens 765 to 832 tok/s, 134,000 tokens 489 to 604 tok/s, 262,000 tokens 335 to 445 tok/s.
LEXI_VK_Q41_TILE=v1restores the previous kernel. - Engine: a rare wrong result in the previous prompt kernel is fixed (a key split shorter than one tile could turn a row into NaN).
-
Engine: prompt working memory is released and reused. Peak video memory during a 3,600-token prompt on the RX 7900 XTX: a 9B model 7.4 to 5.6 GB, Gemma 4 E4B 11.2 to 5.6 GB, same text, same speed.
LEXI_PROMPT_REUSE=0turns it off. - Engine, RTX 2060: a 9B model's prompt processing went from 245 / 233 to 378 / 352 tok/s and its generation from 45.1 / 44.4 / 40.2 to 47.7 / 47.0 / 42.6 tok/s, same text.
- Quantizer calibration runs on every core and accepts recurrent-block projections and projections up to 16,384 wide.
- Against llama.cpp 1.0.0 (released November 13) on one RX 7900 XTX, 27B Q4_K_M at a 262,144-token context with a 4-bit cache and draft-head speculation on both: generation 76.8 against 65.3 tok/s on a short prompt and 70.0 against 55.8 after 33,000 tokens; prompts read at 916 against 725 tok/s (3,600 tokens) and 832 against 605 (33,000). LexiEngine answered 134,000-token and 262,000-token prompts; llama.cpp 1.0.0 returned nothing on either. See Benchmarks.
- LexiOS, LexiKernel and the Panel are as in
5db82d6. - The test sets passed on its source (engine 4,155, GPU server 570, the CUDA kernel suite on the card 163, LexiOS, LexiKernel and Panel as before), and the image passed the nine boots of the VM qualification.
- File:
LexiSuite-5db82d6-secureboot-dev.img.xz(350,878,240 bytes), with its.img.xz.sha256(445e314a4e49ffeaa3a807f4b7db05e9410f25f463c85a594e9c2c13b7a68506) and the decompressed.img.sha256(34b836bd20246ecc6c97f37b0e5f8887045fa19c7106f5388351d639511b0836). Linux entry 7.4-rc2, as in3db31f8. - Superseded by
afaf544above. - Engine: long prompts are read 14% to 37% faster. Every prompt pass rebuilt the row-index lists of its gathers and found the stored copy by comparing the whole list; they are now found by their short per-token form. On a 27B model at a 262,144-token context a 3,600-token prompt went from 660 to 905 tok/s and a 253,000-token prompt from 293 to 336 tok/s, with the same output.
- Engine, Vulkan: prompt attention on the default cache loads each key and value once for the query heads that share it, in three passes whose sums keep their order, so the bytes written are unchanged. The kernel is 9 times faster at 4,096 cached tokens. With default settings a 3,600-token prompt on a 9B model went from 1,400 to 2,877 tok/s.
- Engine, RTX 2060: prompt processing is about three times faster (a 9B model: 89 / 68 to 242 / 220 tok/s, same text): integer Q6_K tiles, compact Q4_K tiles, shared attention scores and kept gather bounds, each with an off switch.
- LexiOS, LexiKernel and the Panel are as in
3db31f8. - The test sets passed on its source (engine 4,154, GPU server 568, the CUDA kernel suite on the card 149, LexiOS, LexiKernel and Panel as before), and the image passed the nine boots of the VM qualification.
- File:
LexiSuite-3db31f8-secureboot-dev.img.xz(350,680,412 bytes), with its.img.xz.sha256(7fb993e68a2d75fe5633081173f0302c67ddf5d956f2f5bd4688c64139c66a1f) and the decompressed.img.sha256(4689a59c544c4ba739d8a0503967f6aea343425af1a548527063096d511341da). Linux entry 7.4-rc2, as in1123d04. -
Superseded by
5db82d6above. This file is on Hugging Face. - Boots LexiKernel on a Ryzen desktop (MSI X570 Tomahawk, Ryzen 9 5900X, 24 CPUs): both NVMe drives and the Realtek RTL8125B 2.5G card bind, the card takes its address over DHCP, serves the Panel to another machine and answers ping, and the USB keyboard works on the local console. Each fix below was found by booting that PC.
- Network card: mapping a device's registers now turns its memory decoding on (firmware leaves it off on devices it did not use, and the card was refused on a garbage revision); the receive ring keeps working past its first 64 frames; a damaged or oversized frame is dropped instead of ending the interface.
- Storage: NVMe 1.2 and 1.3 drives are driven (1.4 was required before). Internal drives stay read-only until approved.
- Local console: USB keyboards and mice are on by default and are read without holding the USB controller; lock LEDs work; typed actions are read (
N,U,Y;RandPwith a typed confirmation in any letter case). -
Setup code:
Lp1-and four groups of four instead of 64 digits; either case, with or without the dashes. - Ping is answered on development media; a refused disk or network card is named on the stop page with its reason; the setup page keeps its hardware-security report for 30 seconds.
- The LexiOS, LexiKernel, Panel and server test sets passed on its source with one test corrected afterwards, and the image passed the VM qualification (the nine boots listed under
1123d04) and a test that types on the local console. The GPU engine test set was last run on4b6d0cdb; no engine code changed since.
- File:
LexiSuite-1123d04-secureboot-dev.img.xz(350,681,260 bytes), with its.img.xz.sha256(8ecf1d43eac73264a889f1ee033c3a1c015f535045c301b859fddbecf555d14d) and the decompressed.img.sha256(a05a9d960444ef024abb2abc220838ce4e0dbea1e0a829ad980aea1e95e45335). Linux entry 7.4-rc2, as in4b6d0cd. - Superseded by
3db31f8above. -
For someone using it, it is
4b6d0cdbelow: the Panel, LexiEngine and LexiOS in it are that source plus one corrected Panel test. Everything listed under4b6d0cdapplies. - New in the source it is built from, and switched off in this image: LexiKernel's own driver takes a physical GeForce RTX 2060 from power-on to a finished job. With the card passed through to LexiKernel in a virtual machine, it runs the card's firmware-security image, the booter and the GSP firmware; the firmware identifies the card (TU106-A, 6 GiB) and lists its ten engines; the instance-memory window works; a GPU address space with host-owned page tables and a command channel on a copy engine are accepted; and a copy-engine job copies a 4 KiB page with all 1,024 words equal. Ten assumptions the synthetic tests had passed were corrected by the card, each with a test. This needs a boot option and a build that carries the vendor's firmware, which release images do not. There is no graphics or compute context yet, so no model runs on this card under LexiKernel.
- The LexiOS, LexiKernel, Panel and server test sets passed on its source, and the image passed the VM qualification: LexiKernel first and second start, the Linux entry, an AMD IOMMU boot, a Ryzen 5000 CPU model, devices behind PCIe ports, and refusal of an unenrolled stick, a tampered rollback file and a tampered kernel. The GPU engine test set was last run on
4b6d0cdb; no engine code changed since.
- File:
LexiSuite-4b6d0cd-secureboot-dev.img.xz(350,681,560 bytes), with its.img.xz.sha256(459b6e3e1a0bb584f30f116fc641ae80333ad3f20f12637ed749dfa6dd6832d1) and the decompressed.img.sha256(7fa13bee353d1d800338c1e06a12ceb3f64769ede03dcc527b2c4788d65f643b). Linux entry 7.4-rc2, as in795b4a9. -
Superseded by
1123d04above. It is on the Releases page as "LexiSuite Testing Image 8"; everything it introduced is in1123d04. - New in LexiEngine: it serves agent applications over the network. Thinking is returned apart from the answer (
reasoning_content); a reasoning level in the request is accepted; tools and tool history go through the model's own chat template and calls are read in the JSON and the function/parameter form; ordinary tool schemas are served; several model names can be declared; a request can wait for a busy slot (--queue-wait); TLS in the engine, or plain HTTP on the network only with an API key. See LexiEngine launch options and API behaviour and LexiEngine on the network. The engine also refuses at start a context that does not fit the card, with the largest that does. - Faults fixed that were found by running an agent application against it: one request with no possible token made the engine refuse every later request; an apostrophe followed by a multi-byte character crashed the tokenizer; a tool call made from the model's thinking was returned as thinking.
- Measured before and after on an idle Radeon RX 7900 XTX: no slowdown on DeepSeek-V2-Lite, a 9B and a 27B model (prompt processing and generation within 1%, the 27B's generation within its run-to-run range).
- New on the stick: a FAT32
LEXISUITEvolume that opens on Windows, macOS and Linux, with the Linux installer, the stick writer, portable Linux binaries and the documents. The Windows and macOS folders hold a note: nothing has been linked or run on those systems yet. The stick cannot yet install itself to an internal disk. - New in the Panel: all 21 languages cover the feature names, tooltips and guided-task text (machine translation, not yet read by native speakers); main can be switched between llama.cpp and LexiEngine; Duplicate copies an instance of any engine. Ternary (1.6-bit) model support is experimental and off unless chosen.
- New in LexiKernel: dynamically linked programs in its ELF loader (tested, not yet used by the booted system);
lk.user_cpus=all; more of its RTX 2060 driver, which on the physical card reads the whole video BIOS and prepares the first firmware step, without yet running firmware on the card. LexiKernel now admits executables up to 128 MiB: the Panel had grown past the earlier 64 MiB limit, which this image's own boot test caught. - The full gate set passed on its source, and the image passed the VM qualification: LexiKernel first and second start, the Linux entry, an AMD IOMMU boot, a Ryzen 5000 CPU model, devices behind PCIe ports, and refusal of an unenrolled stick, a tampered rollback file and a tampered kernel.
- File:
LexiSuite-795b4a9-secureboot-dev.img.xz(321,342,828 bytes), with its.img.xz.sha256(3d2263196b4b7bf391aee3744969b3a488092ba0e1bfb9ac66887f95a43143e1) and the decompressed.img.sha256(ccda3083a7153465cecdcee6424cd1b9dc2f862b3fe0845995563b86675582d3). Linux entry 7.4-rc2, as in5f859a7. -
LexiEngine is unchanged from
5f859a7, so the Benchmarks apply as measured. - New in LexiKernel: its own RDNA3 driver starts a physical Radeon RX 7900 XTX as far as running a command buffer from a user address space with an interrupt-delivered fence (tested with the card passed through to LexiKernel in a virtual machine; no shader yet, so GPU work under LexiKernel still uses the Linux driver-domain guest); a first read-only contact with a GeForce RTX 2060;
TCP_NODELAYin its TCP stack. The GPU bring-up runs only when asked for with thelk.gpu_probeboot option. - The full gate set passed on its source commit, and the image passed the VM qualification: LexiKernel first and second start, the Linux entry, an AMD IOMMU boot, a Ryzen 5000 CPU model, devices behind PCIe ports, and refusal of an unenrolled stick, a tampered rollback file and a tampered kernel.
- File:
LexiSuite-5f859a7-secureboot-dev.img.xz, with its.img.xz.sha256and the decompressed.img.sha256. Linux entry 7.4-rc2, as in0839e30. - New in the engine since
0839e30: draft-head speculation for sampled requests as well as greedy ones; contexts up to 1,048,576 tokens accepted, with admission that charges working memory once; no fixed memory budget (the engine sizes itself from the card); DeepSeek-V2 on Vulkan rebuilt (batched prompt passes, matrix-core attention and experts, 16-bit and latent caches, the previous request's prefix kept). - The full gate set passed on its source commit, and the image passed the VM qualification: LexiKernel first and second start, rollback variable creation, the Linux entry, an AMD IOMMU boot, a Ryzen 5000 CPU model without PCID, devices behind PCIe ports, and refusal of an unenrolled stick, a tampered rollback file and a tampered kernel.
- Supersedes
0839e30. Install withinstall-to-usb.shor an image writer: Install from USB.
- File:
LexiSuite-0839e30-secureboot-dev.img.xz, with its.img.xz.sha256and the decompressed.img.sha256. Same LexiSuite code asc4ce6d8(the classic Panel with tabs on top, drop-down menus and sub-tabs); the Linux entry moves from 7.4-rc1 to 7.4-rc2 (7.4.0-rc2-lexios-usb). - The kernel is built from the
v7.4-rc2tag, whose signature by Linus Torvalds was verified, with the same LexiOS configuration as 7.4-rc1 (no setting changed) and all 133 modules signed. - Passed the VM qualification: LexiKernel first and second start, rollback variable creation, the Linux entry, an AMD IOMMU boot, a Ryzen 5000 CPU model without PCID, devices behind PCIe ports, and refusal of an unenrolled stick, a tampered rollback file and a tampered kernel. Not yet booted on the Ryzen desktop.
- Supersedes
c4ce6d8.
- File:
LexiSuite-c4ce6d8-secureboot-dev.img.xz, with its.img.xz.sha256and the decompressed.img.sha256. Passed the VM qualification: LexiKernel first and second start, rollback variable creation, the Linux entry, an AMD IOMMU boot, a Ryzen 5000 CPU model without PCID, devices behind PCIe ports, and refusal of an unenrolled stick, a tampered rollback file and a tampered kernel. - New since
1c5fa3c: the classic Panel page is the Panel again, with tabs on top, drop-down menus and sub-tabs (Panel Navigation and Help). Pages that existed only in the 11/07 interface open inside the classic frame, and old addresses of pages that only repeated a classic tab open the classic tab. - The full gate set passed on its source commit. Not yet booted on the Ryzen desktop where the earlier images stopped.
- Test image of the 11/08 source, not a published release. Passed the VM qualification: LexiKernel first and second start, rollback variable creation, the Linux entry, an AMD IOMMU boot, a Ryzen 5000 CPU model without PCID, devices behind PCIe ports, and refusal of an unenrolled stick, a tampered rollback file and a tampered kernel.
- New since
051033e: on a PC that does not verify the stick, a development image asks before starting LexiKernel unverified and says NOT VERIFIED on its status page; LexiKernel runs on CPUs without PCID; devices behind PCIe ports (USB and network on Ryzen desktops) get DMA; a stop page shows its diagnostic lines; IOMMU events of every type are reported; the Panel no longer restarts when it records an event on LexiKernel, and threads of a process stay on one CPU again. - Not yet booted on the Ryzen desktop where the earlier images stopped.
- File:
LexiSuite-051033e-secureboot-dev.img.xz; published under theLexiSuitet6release entry. Verify with its matching.img.xz.sha256; decompressed.img.sha256is separately provided. - Boot menu: 1 LexiKernel or 2 LexiOS on Linux 7.4 (actual kernel 7.4-rc1). Both are development-signed with the LexiSuite owner certificate; Secure Boot requires owner certificate append to
db. - Verified in a virtual machine from unenrolled Secure Boot: both boot paths, LexiKernel first-boot anti-rollback-variable creation and reuse, AMD-IOMMU path and refusals of unenrolled/tampered kernel/tampered rollback material. Physical Secure Boot boot remains unverified.
- The
LEXIDATApartition/filesystem grows to fill the USB stick at first boot; this was later confirmed on a physical stick as well as in a VM. - Browser audits fixed issues in System, storage, model upload, start-at-boot, engine-start errors, chat initialization and Secure Boot status. Known LexiKernel limits include CPU/storage measurement and one intermittent freeze after console appearance.
- Development appliance ships no models; CPU/Vulkan upstream runtime on LexiOS passed referenced CPU/GPU text parity tests, with HTTPS downloads validating certificates.
- The native LexiKernel Navi 31 GPU driver remains software/mock-tested only; production and physical hardware GPU qualification not claimed. Read the full status.
- Booted LexiKernel and LexiOS Linux 7.3 from one image. A 32 GiB virtual disk expanded its data partition successfully on both kernels; at the time physical-stick expansion was not yet verified (the later
051033estatus above supersedes this gap). - Signed with a development owner key; initial instructions warned that an unenrolled key would prevent Secure Boot LexiKernel startup.
- Physical Intel laptop boot brought up networking, the console and Panel. Native AMD GPU remained untested; the driver-domain guest provided GPU access under LexiKernel.
- Performance claims for the earlier engine build were superseded by the later benchmark record; do not treat the Linux 7.3 image results as Linux 7.4 image results.
- Legacy checksum label
SHA256SUMSbelongs to this earlier image's release entry, not automatically to051033e.
The following are dated October 4–6, 2026 in the source changelog. They describe changes in Panel/engine and release packaging, not additional tested USB images.
- Integrated 24 performance batches across Panel, engine and OS integration; Responses API composition, GPU residency and scheduling interfaces, benchmark components and capability-guided controls.
- Fixed subset-logits handling, optimizer restoration/promotion safeguards, secret handling and packaging integration.
- Explicit release gates still open: accelerator factories, full GPU graph dispatch and physical-appliance qualification; unavailable operations must fail explicitly.
- Added bundled CPU reference runtime, engine catalog and validated separate-instance launch, with before/after binary backup.
- Added model-share cache, platform planning, runtime capability probes, maintenance windows and guarded integration/test coverage.
- Public qualification remained gated on physical inference, model parity and soak evidence; candidate is not a general-availability declaration.
- Notifications, backup/diff and restore, model library integrity and cleanup, fleet model synchronization, maintenance windows, signed LexiOS evidence and network-share storage locations.
- Added scoped expiring/rotatable tokens and route-derived API discovery, observability, measured tests and configuration-transfer tooling.
- Added guarded network/storage operations with rollback planning and explicit privileged controls.
- Reworked network changes around reviewed apply / confirm / revert workflows; added fleet link measurements and crash evidence.
- Added read-only/native Rust views for Status, Bench, Optimize, GPU/Power, Models, Builds, Files, Access/Fleet and related sections; privileged operations kept their confirmations.
- Added per-instance metrics, bounded telemetry and appliance protocol-v0 integration with explicit Linux fallbacks.
- Reunified the classic Panel as the normal UI, retaining advanced native views at
/rust. - Exposed alerts, engine capabilities, benchmark comparisons and per-instance engine flag checking in familiar tabs.
- Kept Fleet hidden in the default single-server layout unless enabled in Panel layout; customized UI files remain respected.
- Introduced reproducible CPU perplexity/KLD and deterministic compatible-API task evaluation, bootstrap confidence intervals, qualification records and Quant Forge quality gates.
- Added evaluation history/UI and repaired navigation and release-route metadata.
- Added inference-fabric streaming with backpressure, disconnect cancellation for streaming, failover and timing metrics.
- Added engine capability catalogs and validation, node telemetry/alerts, the hardware safety governor, measured tuning and Quant Forge changes.
- Expanded native Panel pages, signed update bundles and localized navigation with English fallback; selective localization was not full-product translation coverage.
- Corrected file/folder download naming, including international filename encoding.
- Added typed decision, scoring, routing and calibration work, including exact model/runtime/candidate binding and later deployment quality gates.
- Added language selection and the earlier inference integration; validation results and restrictions are feature-specific rather than broad guarantees.
- The changelog does not enumerate separate patch releases inside this grouped section; no imaginary version history is implied.
- Added durable bounded inference-fabric state, failover, quantization integrity checks and transaction recovery.
- Hardened authentication, malformed-request rejection, secret comparison and single-user listener restrictions.
- Unreleased does not mean included in every distributed image; inspect the exact build before relying on a capability.
- November 6–7, newer source benchmark record: 27B RX 7900 XTX on host Linux 7.3, not the USB image, controlled context and power; see Benchmarks. The benchmark-context engine revision and later README revision are not identical; compare only runs with known provenance.
- The source snapshot reports approximately 1.78 million Rust lines / 327 crates / 15,405 test functions. Source counts do not establish production readiness.
- LexiKernel still has incomplete CPU/storage telemetry and native RDNA3 physical qualification; Kernels, Hardware, Security explain boundaries.
| Control / artifact | Meaning |
|---|---|
| Image filename and matching SHA-256 | Binds a specific archive to its published digest; never mix checksum files across releases. |
| Boot menu 1 / 2 | Chooses LexiKernel or this image's Linux 7.4-rc1 path. |
Secure Boot db Append |
Adds trust for the matching development signing certificate; future keys need separate enrollment. |
| Storage / Backup / Updates | Review, sign, stage, confirm and recover appliance changes; do not skip verification or rollback policy. |
| Panel version/build display | Helps distinguish changelog claims from the binary actually running. |
Home · Install from USB · Secure Boot · Benchmarks · Licence
LexiSuite · Local OS / Kernel / Engine / Panel · Development / testing; not production-qualified
Getting Started · Security · Troubleshooting · Release Notes · Licence
Verify release checksums, finish first-boot enrollment on a trusted network, and respect the current proprietary licence.
- Home
- Getting Started
- Install from USB
- Write a stick
- Install to an internal disk
- Install on a computer you already use: Linux, macOS, Windows
- Updating
- Ternary models (experimental)
- First Boot
- Secure Boot
- Kernels
- Hardware
- Overview / Home
- Navigation and Help
- Models
- Instances
- LexiChat
- Builds and Engines
- Optimizer
- Fit Lab
- Benchmarks and Evals
- Training
- Decisions
- TokenGate
- Inference Fabric
- Media Generation
- Agents and MCP
- Power and Hardware
- Network and Access
- Storage, Backup and Updates
- Files, Terminal and Passwords
- Diagnostics
- Fleet
- Virtual Machines
- LexiOS and Secure Boot
- macOS