Skip to content

Draft: what ships in v2026.43 - #6

Draft
bong-water-water-bong wants to merge 2 commits into
mainfrom
1bit/post-v2026.43
Draft

bong-water-water-bong wants to merge 2 commits into
mainfrom
1bit/post-v2026.43

Conversation

@bong-water-water-bong

Copy link
Copy Markdown
Contributor

Draft post for the v2026.43 release: the candidate's data, no claims beyond what the measurement database and the proofs carry. Branch kept unmerged — the owner publishes.

What it says

  • Pins the candidate is built from (llama.cpp 851fa6c8d8f2, hrx-system 82804197bcbe, Lemonade 27abcc24da99, our HIP kernels abaa6bc484d5, the sealed NPU kernel sets 17ed5be7a28a), linking to the full list in docs/releases.md.
  • The week's HRX speed, each row with its number: verify step for 2–8 tokens (27B MTP 25.9/20.5/21.9 → 36.0/26.8/29.9 tok/s, step 121 → 88 ms), split-K decode attention (Qwen3.5-4B 32K 23.7 → 47.1 tok/s), Flash-Next expert formats (5.1 → 26 tok/s), prompt FA causal block skipping (Qwen3-4B pp4096 +45 %), prefill tiles (+1.6…2.5 %), the load warm-up (27B short-chat TTFT 2.64 → 0.80 s).
  • Eight concurrent chats (ours standard, 90 W cap, PCM pad (23 Oct)): 27B 50.7, Coder-30B 138.1, Qwen3.5-4B 194.3 aggregate tok/s at N=8 — with the honest comparison that Lemonade's Vulkan build reaches 201.8 and 245.7 in the same run, so the MoE batching gap is named as the next speed row rather than hidden.
  • Correctness: strict placement by default, census 43/43, #315 resolved (37 arms, 9 h 03 min, 0 faults), the ZAYA1-8B NaN report closed on 120 loads, and the gate's new coverage.
  • NPU: sealed kernel sets only, gate phase 90.33 tok/s on 0.6B, sensors and NPU proven from the booted OS image.
  • Lemonade + 1bit OS: one hop, warm-ups in order, 22 HRX catalog entries proven (43.4–330.7 tok/s chat-short) plus 3 NPU entries; the image's MOK-signed UKI and the USB4/TB stack.
  • What is not in it: the OS install step (on a branch) and Windows.

Rules followed

Tone is the site's: what was built and how it measured, no comparisons framed as wins, no vendor commentary. Closed-source wording only — "our HIP kernels", "the sealed NPU kernel sets" — no kernel names, files or methods. No image: line, so the post uses the site's own preview card instead of borrowing another post's.

Verification

Front matter matches the format the other posts use (tags, summary, then # Title); the site build happens in pages.yml with the engine checkout, so this branch was not built locally. Every number is traceable to the auto-opt database rows, the catalog proofs page, a closed issue, or a linked engine PR.

🤖 Generated with Claude Code

bong-water-water-bong and others added 2 commits October 23, 2026 07:45
The release candidate's data as a post: the pins, the week's HRX speed with the measured numbers,
the 8-concurrent-chat cells (ours standard, 90 W cap, PCM pad (23 Oct)), the correctness items
(strict placement, 43/43 census, #315 resolved, the ZAYA1-8B NaN close), the NPU gate at 90.33 tok/s,
Lemonade's one hop with 22 proven catalog entries, and the 1bit OS image. It also says what is not in
the candidate: the install step, and Windows.

Every number carries its standard (lemonade bench standard or ours) and links to docs/releases.md,
docs/hw-gate.md, docs/npu.md, docs/serve.md.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…4bd713d968)

Keeps the post's pin table identical to the engine docs' candidate section, which is what the
release run will record.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant