Skip to content

1bit OS

bong-water-water-bong edited this page Nov 8, 2026 · 9 revisions

1bit OS

1bit OS is a Puppy Linux appliance. It is ResolutePup64, built with woof-CE on the Ubuntu 26.04 "resolute" base, with Lemonade's server and the 1bit engine embedded in it, running our kernel rather than the distribution's. The same image is the live/try stick and the installed system.

what
base ResolutePup64 (woof-CE on the Ubuntu 26.04 resolute base); live from the stick, or a frugal install on a disk
kernel 7.4.0-rc1-1bit, ours, MOK-signed, with amdxdna (NPU, the 0011v2 build, srcversion EE1F36A2B9BA294CE11AD80 on the c26 image) and ec_su_axb35 (fan and EC) signed the same way
Secure Boot on: shim → our MOK-signed grub → our MOK-signed kernel → our signed modules
boot headless by default: a console login on tty1, no X. A second boot entry, 1bit OS (desktop), starts Puppy's JWM desktop on the same system
userland bash, sudo, apt, sshd, mokutil: an ordinary Linux box, not an initramfs
persistence everything that has to last is on the 1BIT-DATA partition, which is Puppy's save layer (PUPMODE=12): models, Lemonade's config and cache, SSH keys
what it serves Lemonade Server on :13305 with its web app, its model catalogue and the engine as its onebit backend: the Radeon 8060S iGPU through HRX and our HIP kernels, the XDNA 2 NPU through our own runtime

Where it stands (2026-11-05)

The appliance boots on the Max+ 395 with Secure Boot on (first on 2026-11-02, cycle 12, after an earlier attempt stopped at a boot-media race that cycle 11 pins to pupsfs=1BIT-DATA and guards with an init that cannot exit). The last results measured on the board are cycle 27's (2026-11-05) and, before them, cycle 24's:

  • puppy-check 42 rows, 1 FAIL, 2 SKIP — the FAIL was the checker's own defect, and the two SKIPs are the expected headless ones; cycle 25's board rows are unrun, and cycle 26's software gate on the shipping artifact is vm-check 75 PASS / 0 FAIL / 3 SKIP (UEFI + Secure Boot)
  • cycle 26's board boot ran on 2026-11-05 and failed before userspace, and the cause is the medium rather than the cycle: init printed Waiting for partition [1BIT-DATA]............... failed. Puppy's wait_for_blkid() polls blkid 15 × 1 s and gave up at 15 s, while this board's stick — behind two resets of the JetKVM's own emulated USB mass-storage interface on the same xHCI controller — only bound [sda] at 45.7 s (its /data/init-fail.log). c26's vmlinuz, initrd.gz and grub.cfg are byte-identical to c25's, so this is not a c26 regression, and every VM row was green because QEMU publishes the partition at ~3.7 s (engine#1159). Cycle 27 re-cuts c26 with one file changed, /1bit/initrd.gz: the wait is 120 s and a still waiting line prints at 15 s, and engine#1160 makes the image build refuse an initrd that waits under 90 s or waits silently. c27's gates on its frozen bytes: vm-check 75 PASS / 0 FAIL / 3 SKIP, usbgate.sh PASS (the new row, which asserts that budget out of the shipped initrd and then boots it on USB mass storage), prepare-stick.sh --check PASS, and the write READBACK: MATCH over all 9,128,902,656 bytes.
  • c27 booted this board on 2026-11-05 — sda1 sda2 at 3.86 s (the same slow-medium shape that killed c26), boot to Lemonade serving in 50 s — and its puppy-check measured 51 rows, 43 green, 1 FAIL, 7 SKIP. The FAIL is u4-release-check : no RESULT line; rc=1: the release check died at line 1244 on PPID=$!, a readonly bash builtin, which left its 1000-token chat in the NPU driver, so the next backend's requests came back code 102 behind 33 aie2_hwctx_init … ret -19 lines and one TDR followed at kernel 319.2 s (engine#1177). Before that death the run was fully green: NPU chat 94.26 tok/s, chat-npu 98.27, chat-hrx 82.83, a 512-token follow-up with TDRs 0 → 0 → 0 → 0, a 296-chunk stream after evict and reload, the 400/600/1000 ladder answered and the queue row PASS; the demo rehearsal on that boot was READY (RAG 3/3, game 5/5, picture). The five u9-* SKIPs were the checker's own path bug (/data/logs/<stamp>-<bootid> against /data/logs/<bootid>), fixed the same night
  • cycle 28 is c27's bytes with that one fix (1bit-release-check 5b2f5cc0… for 729e5a1d…; kernel, zdrv and /1bit/initrd.gz byte-identical) and booted the board at 19:42 on 5 Nov: puppy-check 52 rows: 47 PASS, 3 FAIL, 2 SKIP, boot to Lemonade serving in 51.5 s, NPU 94.9–98.0 and HRX 85.4–87.1 tok/s (its VM usbbootgate had already passed, including the image's own release check RESULT: PASS (25 checks, no failures)). The three reds are a non-streamed NPU chat the release check saw get no HTTP status, that check's own follow-up run on the same wedge (code 102, one TDR), and u9-degraded's first real firing. cycle 29 composes c28 with the console and installer fixes (1bit-install's re-install chain, 1bit-selftest now shipped, the Install menu row, extra_models_dir) and is cut but not released; it booted the board at 22:13 for the demo take (one code 614 NPU turn, recovered to the fallback), which is not a puppy-check run, so a full row set on c29 is still owed
  • no Lemonade API key: since cycle 20 the appliance ships with a blank key, so Lemonade on 0.0.0.0:13305 answers like stock Lemonade, and console menu item 1 mints one and prints it
  • the catalogue carries the staged 1bit model — cycle 26 stages Qwen3-4B-Instruct-2507-1bit on the onebit backend; cycle 12's board measured chat 200 OK, 70.49 tok/s, TTFT 1.30 s with 0 amdxdna errors

In QEMU, on the same build: G2 (boots on our kernel, layers its SFS), G3 (amdxdna and ec_su_axb35 present, signed, verify against the enrolled MOK), G7 (the save layer persists across a reboot, PUPMODE=12 on 1BIT-DATA), the init guard (panic-repro.sh, with the boot pin and with it deliberately broken) and vm-check with no failing row and Secure Boot on.

Nothing is published yet. The next release carries the appliance: v2026.44 is a draft, and the 1bit-os-* artifacts of the c18-c20 candidates are the BusyBox os/mini image, which is a CI/VM target and not 1bit OS.

When it is published, the image comes from Hugging Face, not from the GitHub release: 1bit-os-puppy.img.xz (about 2.2 GB compressed, 9,128,902,656 B written) with its SHA256SUMS, because a GitHub release asset is capped at 2 GiB. The compressed size is not quoted exactly on purpose: every cut changes it, and the figure to check is the one ls -l prints next to the file you downloaded, not one from a page. That repository is private until v2026.44 is published, so it answers 401 before then. Write it with

sha256sum -c SHA256SUMS --ignore-missing
xz -dc 1bit-os-puppy.img.xz | sudo dd of=/dev/sdX bs=4M conv=fsync status=progress

(or point balenaEtcher / Rufus in DD mode at the .img.xz, which they take directly) onto a stick of 16 GB or more. Writing it with dd alone leaves the stick with no Lemonade API key, so also put the key and your SSH key on 1BIT-DATA — the write and prepare steps, and the Secure Boot notes, are docs/install.md (engine#867). The GitHub releases page keeps the lemonade-onebit package, the release notes and the release SHA256SUMS, and links there for the image.

The three bases in the engine tree

where what it is shipped
os/v3/puppy/ 1bit OS, the appliance above: the woof-CE recipe, our kernel as woof-CE's kernel package, the probes and the gates the next release
os/v2/ the same appliance on a minimal Ubuntu 26.04 root filesystem with systemd, updated with .debs from 1bit.gg/apt; the documented fallback no
os/mini/ the BusyBox image: one EFI file, everything in RAM. Not 1bit OS: the QEMU vmtest/vmtest-release target, and the medium the c18-c20 candidates carried no

Why Puppy and not our initramfs (the short version)

The initramfs cannot grow a package manager or a persistent layer: everything is in RAM and the update path is a reflash. Puppy brings both, plus a save-layer model (pupmode 5/12/13, save file or folder) that is the answer to engine#706, at the cost of size (1.25 GB ISO, 1.3 GB unpacked) and boot time (108.8 s to the first screen in a 2 vCPU / 2 GiB QEMU, against 24.5 s for the initramfs). Our kernel plus our drivers, not the distribution's, was the part that had to be built: a kernel config gap of about fifteen =y symbols (os/v3/puppy/kernel-puppy.fragment), not a fork. The measurements, the four blockers and the gates table are in docs/os-v3-live-puppy-eval.md; the design is docs/os-v3.md.

What is still the board's, and not claimed

  • the install to the board's own internal disk, and the boot from it with the stick out (QEMU has done both)
  • the HRX catalogue in the appliance: 1 model today, where the c20 image listed 23 HRX + 4 NPU onebit entries
  • the NPU through Lemonade, the desktop on the JetKVM console, and the device rows of puppy-check

Clone this wiki locally