[iamwavecut]

What's New

The README gets a public changelog under its description: one dated line per delivered milestone,
newest first, with links to the PRs, model cards, docs and the Telegram announcements. It covers
2026-09-23 (Ternary Bonsai 2 on one RTX 3090) to 2026-10-25 (DeepSeek-V4.1-Flash, tensor
parallelism), including pipelines across GPU architectures (#92, #93) and CPU+GPU expert
execution (#103, #108).

DeepSeek-V4.1-Flash in the README and docs

• README: the artifact row with its Engram table, the formats sentence, the 24 GB row of the memory table.

• docs/recipes.md: download and serve commands with host and disk experts; the flag table names
  Flint and DeepSeek where they apply.

• docs/gpus.md: the 24 GB row and the RTX 3090's models.

• docs/maintainer/pipeline-stages.md: tensor parallelism refuses DeepSeek too.

• docs/maintainer/engine-architecture.md: the Gemma 4, Flint and DeepSeek families and their
  Engine cores in the model text and the ownership table.

• CONTRIBUTORS.md: the DeepSeek artifacts' sources and references.

• docs/recipes.md also gets a Flint section.

DeepSeek-V4.1-Flash on the RTX 4090, 5090 and PRO 6000

Short runs of the native release build of b3402a6fd (CUDA 13.4 Linux archive) on rented Vast
hosts, one GPU each, the published model and table, the 14-token and 8,086-token prompts of the
RTX 3090 runs, host experts and then disk experts with the file in the page cache:

Card	Experts	Decode, short	after 8,086 tokens	Prefill of 8,086	Ready
RTX PRO 6000	host	48.0	45.3	1,310	64 s
RTX PRO 6000	disk	37.4	33.4	431	29 s
RTX 5090	host	29.2	22.5	499	34 s
RTX 5090	disk	12.5	10.5	181	15 s
RTX 4090	host	25.9	22.2	490	41 s
RTX 4090	disk	8.5	7.3	123	14 s

Host and disk produced the same tokens on each card; the short answer's length differs between
cards (72, 81, 78 tokens; 82 on the RTX 3090). Added to docs/deepseek-v41.md,
docs/performance.md (and its Cards table), the README Highlights and What's New, and the cards'
rows in docs/gpus.md. DeepSeek runs on one GPU only, so there is no two-GPU NInfer row; the
RTX 3090 host's Strata rows keep its one- and two-GPU numbers. Each row is one run.

The archive links FFmpeg 7.1 (libavformat61); the rented hosts ran Ubuntu 26.04 (FFmpeg 8), so
the four libraries came from Debian 13.

The version is the release tag

• cmake/GenerateBuildId.cmake: NINFER_VERSION from the environment, else the release tag
  (YYYY-MM-DD.N) at or before HEAD, else untagged. Products keep <version>+<build id>.
• native-release.yml: a run that will publish allocates its tag in validate and passes it to the
  Linux and Windows builds and to publish; the job lists draft releases, so it gets
  contents: write.
• native-release-tag.py --tag and publish-native-release.sh: publication checks that the tag
  the binaries report is still this revision's (its draft or release, or a number nothing took
  since) and stops otherwise.
• Packagers take the version from the build's stamped header and refuse one that differs from
  NINFER_VERSION; VERSION and the RELEASE_NOTES_* lookup are removed.

Checks

• python3.11 -m unittest discover -s tests/release: 122 tests pass (new: tag check at
  publication, a taken tag, a retry under another tag, a missing NINFER_VERSION, the package
  version and a mismatching build).
• cmake -P cmake/GenerateBuildId.cmake in this checkout (2026-10-25.1), with
  NINFER_VERSION=2026-10-26.1, and in a repository without tags (untagged).
• actionlint on the three changed workflows; git diff --check.
• Not run: package-release.ps1 (no PowerShell here); its first run is the next Windows release.