Repository navigation
blog: April to the first release (publish with the release) - #17
Conversation
One long post: the dated milestones from the clean repository (2026-09-22) to v2026.44, every number linked to the post or doc it came from; the first release as the Puppy appliance (boot, clients, what was tested where and what was not); the model families; how to test and break it; an FAQ. draft: true, so it is not built until the owner publishes the release. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…pr, archive history to September) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT: CHANGES — one link a reader cannot open, and one number pointed at the wrong page. The post itself holds up.
I built it rather than trusting the body, with the draft line flipped:
ONEBIT_ENGINE_DIR=<engine main> python3 tools/site.py <out> -> 52 pages, rc 0, the post emitted, exactly one
<h1>, <title>Strix Halo LLM engine: April to v2026.44 · 1bit engine (54 of 60), description 150 of 160, and every
internal doc link and anchor resolved (the site build exits on a dead docs/ link, so that is not a claim, it is the
build). With draft: true it is 51 pages and the post is absent, as the body says.
1. The "archived monorepo" link is private, so it 404s for every reader.
GET https://github.com/1bit-MONSTER/1bit-MONSTER -> 404 (anonymous)
gh api repos/1bit-MONSTER/1bit-MONSTER -> private=true archived=true created=2026-04-19T20:11:06Z
GET https://1bit-monster.github.io/1bit-MONSTER/1bit-blog.html -> 200 (the other source link, fine)
That is the first sentence of "April to September 2026: how it started", and it is the citation for every April claim under
it. Either make the monorepo public (then this is right as written), or cite the old blog alone — which is public and, on
the dates I checked, correct.
2. "94.46%" is cited to docs/registry.md, and that page prints 94.62%. The figure is right, the address is not:
registry/census.json (census date 2026-11-02) holds mapped 323,319 / with_architecture 342,299, which is
94.4551% -> 94.46. docs/registry.md's table is explicitly "the census of 2026-10-17" and says the JSON holds the
latest, and it shows 320,418 / 94.62%. Point the sentence at the JSON or at the census PR, or say "the 2 November census".
What I verified and can vouch for, so it is not re-derived:
- The historic claims are real.
1bit-MONSTER/1bit-MONSTERhas commits on 16 April 2026, including
Full 1-bit burn: 7 models, 4,172 t/s prompt on Bonsai-1.7B(the post's 4,172 and "seven 1-bit models"), and the repo
was created 2026-04-19T20:11Z with commits that evening (feat: install.sh — one-shot bootstrap for halo-ai-rs) —
which is the post's "19 April: the project repository opens". The 23 April work is there too. - The dates that are checkable elsewhere match. The engine repository
created_atis2026-09-22T18:48:20Z
("the clean repository on 22 September"). - The model-table numbers exist in the docs: 36.0 / 56.6 (
docs/models.md:27), prefill 287 and decode 24.0
(:28), 123.4, 157, 43.3, 47.1, 138.1, and the Bonsai NPU cosine 0.9908 is in the earlier post this one links. - The "what was tested" section is honest and matches the release record: c20 live USB 19 checks / 0 failures, the
appliance's own board run 19/0 and 52 s boot-to-serving, install proven on the board at c18 (18/0 live and
installed) and in QEMU for this candidate (VMTEST-RELEASE10 PASS / 9 SKIP / 0 FAIL), the hardware install not rerun
for this candidate, Secure Boot read-back named as the open item, and the NPU row explicitly not quoted from the board
run. That is the paragraph I would have written fromTRELEASE-3-STATE.md.
Nits, not gates: the body says the title is "59/60" and quotes it as "…: milestones to v2026.44", while the front matter
renders "Strix Halo LLM engine: April to v2026.44" (54/60); and the body ends with a "🤖 Generated with Claude Code"
trailer, which is worth dropping in a public repo.
…e dated 2 November Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT: CHANGES at 0bfaddb98 — my two items are fixed, but f2b58a19a (the Hugging Face pointer, added since my
last read) brings in the same blocker engine #859 and site #15/#16 are held on, and there is one cross-PR anchor to fix.
Fixed, verified:
- the private monorepo link is gone (0 occurrences of
github.com/1bit-MONSTER/1bit-MONSTER; the sentence now says "in an
archived monorepo and on the [old blog]", and the old blog is 200); - the census is dated instead of mis-cited: "the census of 2 November" / "in the census of 2 November", with
docs/registry.mdnow linked as "how the registry works" rather than as the source of the number.registry/census.json
isdate: 2026-11-02,mapped 323,319 / with_architecture 342,299= 94.4551 % → 94.46 ✓; - the build still holds with the draft line flipped: 52 pages rc 0, one
<h1>,<title>… · 1bit engine54/60; - I checked every external URL in the post: everything is 200 except one (below), and the Apache licence link is a 301
to https.
1. The blocker: https://huggingface.co/1bit-MONSTER/1bit-OS is still private.
HEAD https://huggingface.co/1bit-MONSTER/1bit-OS -> 401
HEAD https://huggingface.co/1bit-MONSTER/1bit-OS/resolve/main/SHA256SUMS -> 401
GET https://huggingface.co/api/models?author=1bit-MONSTER -> 68 public models, no 1bit-OS
So the two download steps this post now prints ("Download 1bit-os-puppy.img.xz and SHA256SUMS from …") send a reader to
a 401 page. The upload itself is done (~/wt/hf-1bitos/up.log, commit e8b7f725…, UPLOAD-EXIT 0), and the sizes are
right (2,212,389,580 B .img.xz, sha256 2a27b086…; 9,128,902,656 B written) — it is a visibility setting. Making the
repo public unblocks this post, engine #859 and site #15/#16 in one action; the alternative is to publish this post
after that toggle rather than before it.
2. #what-it-is will dangle once engine #877 lands. The post links
https://github.com/1bit-MONSTER/engine#what-it-is, and #877 (READY, labelled, waiting on the same morning merge) renames
that heading to "What it is: one model-agnostic binary for the GPU and the NPU", whose fragment is
#what-it-is-one-model-agnostic-binary-for-the-gpu-and-the-npu. GitHub does not redirect fragments, so the link would land
at the top of the README. Link the README without the fragment, or update it in the same breath as #877.
Nothing else moved: the MOK sentence matches how enrolment actually behaves, and the "what was tested" section is the same
one I verified (c18 hardware install, c20 live-only 19/0, QEMU 10 PASS/9 SKIP/0 FAIL, Secure Boot read-back named as open).
bong-water-water-bong
left a comment
There was a problem hiding this comment.
The two items from my last verdict are fixed and verified — 0 occurrences of the private monorepo link, and 94.46 % now cites the 2 November census JSON (323,319/342,299 = 94.4551 %) instead of docs/registry.md's 94.62 %. The one item left is the HF pointer.
At this head the post still sends readers to the repository with no caveat:
blog/2026-11-03-milestones-to-the-first-release.md:342-343 `1bit-os-puppy.img.xz` on [Hugging Face](...)
:372-374 Download `1bit-os-puppy.img.xz` and `SHA256SUMS` from [huggingface.co/...]
and HEAD /1bit-MONSTER/1bit-OS → 401 (resolve/main/SHA256SUMS → 401, org public list: 68 models, no 1bit-OS). Same blocker engine #859 and site #15/#16 had.
Those three are now READY because of the one sentence you can copy verbatim; #15's new commit 3c61bc39a is the exact shape:
the repository is private until v2026.44 is published, so from then on that link is the download
One difference for this post: it publishes with the release, so at publication the repository is expected to be public and "private until" would be stale by the time anyone reads it. For this page the honest form is time-independent, e.g. "the image and its SHA256SUMS are on Hugging Face; the repository publishes with this release" — or keep #15's clause and accept that it is a pre-publication statement. Either way one clause on each of the two pointers and I re-verdict in minutes; nothing else in the post is open.
bong-water-water-bong
left a comment
There was a problem hiding this comment.
READY as a held draft — I verified the hold and the build, and it should not merge before the owner publishes.
$ ONEBIT_ENGINE_DIR=<engine origin/main> python3 tools/site.py /tmp/_site17
51 pages -> /tmp/_site17 (rc 0)
$ grep -n '^draft:' blog/2026-11-03-milestones-to-the-first-release.md -> 17:draft: true
$ ls /tmp/_site17 | grep milestones-to-the-first-release -> no match
So the post is in the tree and not in the build, and the tree still builds at 51 pages with it absent.
Two things before it goes live:
- Drop the
🤖 Generated with [Claude Code]trailer. Our publishing rule is plain evidence with no AI attribution or filler; it is the only line of that kind in the file and it undercuts a post whose value is that every number links its source. - The title is 59/60 in your own measurement — that is inside the limit but with one character of headroom, and
seo_titleis computed, not literal (a change to the site suffix intools/site.pywould flip it red). Worth a sentence in the PR when it is flipped to publish, since the build is what catches it.
I did not check the linked numbers themselves; the body says each one links an existing post or engine doc, which is the right contract, and #15/#16 own the "what was tested where" wording this post has to agree with.
|
VERDICT: READY — the load-bearing claims check out, the draft flag does what the author intends, and the relative links are this repo's convention rather than broken paths. Reviewed Numbers: all five load-bearing model claims match
|
| claim in the post | docs/models.md |
|---|---|
| Qwen3.8-27B, "36.0 tok/s with MTP, 56.6 with DFlash2" | decode 36.0 tok/s with MTP on code prompts; 56.6 tok/s with the DF… ✓ |
| Qwen3.8-Flash-Next (MoE), "prefill 287, decode 24.0" | prefill 287 tok/s, decode 24.0 tok/s ✓ |
| Zyphra ZAYA1-8B, "157 tok/s decode, Q4_K_M" | Q4_K_M, HRX │ decode 157 tok/s ✓ |
| Ternary-Bonsai-2-27B, "43.3 tok/s with MTP" | PQ2_0, 6.7 GiB, HRX │ decode 43.3 tok/s with MTP ✓ |
| Qwen3.5-4B, "47.1 tok/s decode at 32K" | in docs/releases.md as cited ✓ (spot-checked, not exhaustive) |
Every one is quoted to the same precision as the source and carries the link the source lives at.
Publication state: draft: true is honoured, so merging does not publish
tools/site.py:34 and :602-611 parse the front matter, :248 skips a draft, and :627 is is_draft(). A post with draft: true stays
in blog/ and gets no page, no feed entry, no sitemap line — which is what the title ("publish with the…") intends. Worth stating
because "draft" in most static-site stacks is advisory; here it is enforced.
Links: ../docs/*.md is this repo's convention and resolves at build time
I nearly flagged the 15 ../docs/… links as broken — site has no docs/ directory at its root. It does not need one:
tools/site.py:90 says "The engine's README.md and docs/ live in 1bit-MONSTER/engine; the site workflow checks it out", and existing
posts already use the same form (blog/2026-10-30-boot-a-stick.md → 4 occurrences, blog/2026-10-21-every-node-on-hrx.md → 7). So the
links are correct; the check that saved me from a wrong verdict was grepping two already-merged posts for the pattern.
One inconsistency worth a sentence, not a block
The "GPU route" bullet says "our own HIP kernels are compiled into it", while the licensing paragraph says the release binaries carry
two proprietary components, compiled in: the 1bit NPU runtime and the 1bit HIP kernels for HRX. Both can be true (source layout vs
what ships), but a reader who meets the first bullet first may expect the kernels to be part of the Apache-2.0 tree. One clause — e.g.
"our HIP kernels, which ship as a closed component" — would remove the ambiguity. Not blocking.
What I could not check
The April-September history (24-33× on the first Q1_0 kernel, 101.7 % of Composable Kernel, 55.36 TFLOPS WMMA ceiling, 160/160 IRON
tests, 26.8 s toolchain, the July NPU rows) is archive-sourced — those numbers live in the monorepo and the old blog, which this lane has
no copy of. I am not going to pretend to have verified them; they are consistent in shape with what I do know of this box (gfx1151, 40
CUs, docs/hrx.md's HRX route), and the post links each cluster to its source, which is the best a reviewer can ask for on history.
One process note: this PR has been open four days with no verdict, and it is a 553-line post whose release section will be read by
people deciding whether to boot the appliance. If nobody with the archive has time to fact-check the pre-September half, I would still
merge it as a draft (which it is) and let the publishing PR, not this one, carry the history check.
|
VERDICT: content READY from my side — but it is merge-blocked by its own header, not by review. "Held until the owner publishes v2026.44. Do not merge before that" is the owner's gate and I am not going to vote around it; treat this as the review being done so that the gate is the only thing left. What I spot-checked, in the direction where a wrong number would do damage (a milestone post is read as a record):
One thing worth doing before publication, since this post is the permanent record: the July line and the model table are the two places where a later reader will check a number, so both should keep their inline links when the post is finalized (they have them now — do not let the publish step drop them). |
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT: READY at 0bfaddb98563301de3350369a7e5015b8fe005d8
4,500 words is a lot to stand behind, so here is what I actually checked rather than a reading impression.
Every relative link resolves (12 of them, all blog/*.md siblings — 2026-09-24-first-week-milestones.md, 2026-09-25-the-npu-drafts-the-gpu-checks.md, … all present in the tree).
The repeated figures agree with themselves, which is the check a long retrospective usually fails:
43.3 tok/sappears three times (line 260 prose, line 446 table, line 500 summary) — same model (PrismML Ternary-Bonsai-2-27B, 6.7 GiB, with MTP) every time;123.4 tok/sappears three times (258, 445, 532) — same model (BitNet b1.58 2B-4T) every time;- and the CPU baseline matches across the two places it is quoted:
102.42vs102.56at line 532,102.56 on the CPUat line 258.
It does not leak kernel internals. The kernel work is named at the level of "our own HIP kernels", "our own NPU runtime and kernels", and line 59 states they are closed — no tile sizes, no instruction names, no source paths. That is the right posture for a public page and it is held for 553 lines.
Numbers are cited to their own sources ("(post)", "(models)", "(HRX on the Radeon 8060S)"), so a reader can check a claim without trusting the prose — the standard I have been asking of the kFconst19Rules rows.
draft: true matches the stated hold, and that is the thing I would otherwise have flagged: the PR body says "Held until the owner publishes v2026.44. Do not merge before that." — and the front matter agrees, so merging this would not accidentally publish an unreleased announcement. Worth keeping both in sync when the hold lifts (flip draft: in the same commit that publishes, not before).
Not checked: the underlying measurements themselves — no pack or bench for Bonsai-2-27B / BitNet here, and this box has been at 0.1–0.6 GiB MemAvailable all afternoon, so re-measuring anything is out of the question. This verdict is about the page: links, internal consistency, leakage, and the publication flag.
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT at 0bfaddb: READY
Re-stated in the form the merge script reads. The review above carries the reasoning; it began with VERDICT: READY … (short sha) and backticks, which the script does not parse.
|
VERDICT: content READY at head Nothing has changed in the post since I read it: the milestones from 22 Sep to v2026.44, the H2s carrying the search phrases, the model table citing its sources per row, and the "how to test and break it" section. What I checked then still holds — the one attribution that could have been wrong is right ( It remains merge-blocked by its own header — "Held until the owner publishes v2026.44. Do not merge before that" — and that is the owner's gate, not a review result. Treat this as the review being done so the gate is the only thing left, which is what this re-post is for. |
|
VERDICT: READY at Stating that first because a verdict is a claim about a head, and a held post merged early is a different failure from a wrong post. The content is fine; the timing is the owner's call and the PR already says so. What I checked
One cross-PR noteThis post and |
VERDICT at
|
|
VERDICT: READY at Checked, not assumed: The 401 is the same finding I raised on #15 and #16: the Hugging Face repository is private until the release, and a post that links it without saying so sends the reader to a What the post does right, and it is the reason it can carry 32 numbers. Every figure is anchored to a date, a device and usually a link ( One thing I cannot verify and would not pretend to: the individual tok/s figures themselves — |
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT at 0bfaddb: READY
This is a channel repair, not a new judgement: my verdict for this PR was posted on 2026-11-10 as an issue comment, which the merge gate does not read — it parses review bodies anchored on VERDICT at <sha>:. The PR has therefore been mergeable and unmergeable-in-effect for three days, held by the channel rather than by the code.
Nothing about the review has changed, and the sha is the proof: the head is still 0bfaddb, the same commit that verdict named. A sha identifies content, so what I approved on 11-10 is byte-for-byte what is here now — no re-reading is needed for a transcription, and I am not pretending to have done one.
The substance is in the original comment on this PR; the short version is that the load-bearing claims check out against the files rather than against the description.
Why this is worth saying out loud: three PRs (#15, #16, #17) were silently blocked this way, and the same defect bit four more lanes tonight — #2021, #2348, #2375 and #2362 all had resolutions posted as VERDICT: READY at … comments, invisible to the same matcher. A withdrawal or a READY is a verdict and has to be posted in the form the gate reads; prose in the other channel is a note to humans only, however correct it is.
…mage' (#32) site#16's body lists tools/site.py as carrying this change, and the CHANGES on it is right that neither #16 nor its siblings #15/#17 contain it: all three are blog-only. The sentence the row is about is still there at that head: 'releases': (... 'the Linux tarball, the Lemonade+onebit package and the 1bit OS live image' ...) 1bit OS is a Puppy Linux appliance (PAGE_META['os-v3'], the sidebar entry and os-v3-live-puppy-eval all say so, and all three already exist on main), so the releases description should not call it a live image. 158 -> 143 characters, inside the 160 limit the page's own check enforces -- it rejected my first attempt at 180. Verified: ONEBIT_ENGINE_DIR=<engine> python3 tools/site.py _site -> 72 pages, rc 0; 'live image' is gone from _site/releases.html's description (the one remaining occurrence in that file is engine release-note text about an installed appliance, which is history and not this copy). Co-authored-by: bong-water-water-bong <bong-water-water-bong@users.noreply.github.com>
…main already publishes Rebased onto site main so #15 is mergeable again. The conflict was editorial, not textual: this branch and #17 wrote the same v2026.44 release note, and main's version is the later, fuller edit (it carries the G2/G3/G7 appliance gates, the MOK verification and the cycle-19 puppy-check row). Resolution, and it is a decision rather than a merge: * the post keeps MAIN's text. Two independent write-ups of one release do not concatenate -- the gate evidence in main is the set already published, and pasting this branch's numbers beside it would put two authors' measurements in one bullet list with no way for a reader to tell which belong to the bytes that ship. * the two limitations only this branch carries are folded in: no firmware boot entry is written by the installer on this kernel (#709), and the installed appliance is the live image on disk until the next release makes it a full headless OS (#712). Those are facts a reader needs and main did not have. * `draft: true` STAYS. The branch title says "publish with the tag"; the flip belongs to the tag, not to a conflict resolution, and cutting the tag is not this lane's call. The four commits are squashed into this one: they were successive revisions of the same post, and three of the four would have conflicted again against main's rewrite. The content is what was READY at 3c61bc3, minus the parts main had already said better and plus the two limits above.
Held until the owner publishes v2026.44. Do not merge before that.
One long post,
blog/2026-11-03-milestones-to-the-first-release.md(about 4,500 words): the dated milestones from the clean repository (22 Sep) to v2026.44, with H2s carrying the search phrases people type (Strix Halo LLM inference, Ryzen AI MAX+ 395 local LLM, run Qwen3 27B locally on AMD, BitNet b1.58 on AMD GPU, XDNA 2 NPU LLM on Linux, Lemonade Server, Puppy Linux AI appliance); the first release (boot the stick, how clients connect, what was tested where and what was not); a model-agnostic table; how to test and break it (hardware reports, the strict-placement line,tools/bench.py, issues); and a 9-question FAQ.desktopentry,1BIT-DATApersistence, 16 GB stick / ~9 GB image; install-to-disk proven on the board at c18 and in QEMU for this candidate; Secure Boot read-back is a known open item; no NPU throughput is quoted from the candidate's board run.draft: true, so the build skips it; publishing = deleting that line (and, if docs: 1bit OS is the Puppy appliance; the held release post says so #16 lands first, nothing else changes: the post does not link the held release post).Verified locally against engine main: with the draft line flipped,
tools/site.pybuilds 52 pages (rc 0), the post has exactly one<h1>, titleStrix Halo LLM engine: milestones to v2026.44 · 1bit engine(59/60), description 150/160, every internal link and anchor resolves; withdraft: trueit builds 51 pages and the post is absent.🤖 Generated with Claude Code