[Bizuayeu]

What this changes

One row in the README's Pick a model, after GLM-5.3-Flash Jammed XSFP4, and one sentence in the GLM CUDA recipe's opening. #2894 reads NVFP4_NZ5 checkpoints, and Jammed XSFP4 NZ5 is the checkpoint in that form. The row says "on main after 1.2.7", which holds once #2894 is in, so this one lands after it. The row is shorter than its neighbours and leaves the numbers to the card it links.

• Issue: Refs glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894, which reads this checkpoint's form.
• Stacked on: main (000b534a28). Needs glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894 merged first: until then, the row and the recipe sentence describe code that is not on main. With glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894's head, this merges clean.
• atc overlap --json: complete: false (otherPullRequestsLimit, pullFilesUnavailable, timeLimit). Of the pull requests it read, flint: serve Recursive Flint (40B-A2.5B MoE) natively on Metal #1177, lfm3: serve LFM3-1.7B on CUDA: NVFP4 and FP8 weight-only, MLX 4-bit (stacked on #1513, #1492 and #1503) #1508, lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508) #1634, Implement GLP control vectors with a fused Metal kernel for Qwen3.8-27B #1835, glm cuda: TF_GLM_GRAPHS=1 replays verify windows as CUDA graphs (#1490) #2279 and cuda: a shared engine for dense decoders, with dots.mocr's text decoder on it #2673 share a file with this one. flint: serve Recursive Flint (40B-A2.5B MoE) natively on Metal #1177, Implement GLP control vectors with a fused Metal kernel for Qwen3.8-27B #1835, glm cuda: TF_GLM_GRAPHS=1 replays verify windows as CUDA graphs (#1490) #2279 and cuda: a shared engine for dense decoders, with dots.mocr's text decoder on it #2673 merge clean; lfm3: serve LFM3-1.7B on CUDA: NVFP4 and FP8 weight-only, MLX 4-bit (stacked on #1513, #1492 and #1503) #1508 and lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508) #1634 already conflict with main in README.md.

Checks on this head

• Head commit: f048007945
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: n/a, the diff is README.md and docs/recipes/glm-5.3-flash-cuda.md, no source. python3 tools/pr_check.py --base origin/main: 2 changed files, lean 345 vs 345, 0 new, 0 errors and 0 warnings.
• Tests run, with the command and pass, fail and skip counts: n/a, documentation only.
• CI: on this head, all 12 jobs that ran passed (fmt, zig, zig-windows, zig-macos-dev-tools, pytest, release-paths, both benchmark-clients, desktop, desktop-linux, appimage-linux and mobile); host-nightly-linux and host-nightly-macos were skipped, as they are on pull requests.
• Not run, and why: the inference checks, as the change cannot reach inference.

Receipt

• n/a: documentation only. What the row says comes from these:
  • Served at TP=2: its card's In serving, on two GB10s with TensorFold's main at fde40f638b and glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894's change, a 40,960-token window, one stream and vision off: rank 0's model 83.53 GiB, and three fixed prompts of 512 greedy tokens each gave the same token_sha256 drafted and with "draft": false. glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894's own receipts were taken on its head.
  • Served at TP=2 on main with glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894: as glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894 is not in yet, main at 3885dceb2b with glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894's head 39abe29739 merged into it (no conflicts), built with an empty global cache, Zig 0.17.0 and nvcc 13.3 for sm_121. Two GB10s over a direct link (driver 580.178.04), Bizuayeu/GLM-5.3-Flash-Jammed-XSFP4-NZ5@1c900ea0 from a local copy of the same bytes, a 40,960-token window, one stream and vision off. The first serve after the page cache was dropped, with the checkpoint's digest read in full (0 of 132 files cached): loaded in 109.7s and rank 0's model 83.53 GiB. The same three fixed prompts, 512 greedy tokens each, gave the same token_sha256 drafted and with "draft": false, 8644311e2865, c573bfb5a552 and 72bbf00e4cb0, as on glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894's b4ced9901e. Once glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5 #2894 is in, I will serve it again on main and replace these if anything differs.
  • Its size: the safetensors files at the pinned revision, 160.17 GiB, the same as Jammed XSFP4's (0a22bb46).
  • Its closeness: its card's Closeness to BF16, the same 128 documents in PyTorch for both checkpoints: KL 0.0440 against Jammed XSFP4's 0.0470, top-1 agreement 93.66% against 93.40%, and NLL 1.48% above BF16's against 1.72%. Each of the three differences has a 95% interval, from a paired bootstrap over the documents, that excludes zero.
  • Licence: no weights in this change; the card gives the checkpoint's licence.

Checklist

• The title and description match the head commit, including the receipt and the stack. (f048007945, one commit on main at 000b534a28.)
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No tests: documentation only.)
• Comments and docstrings are one line. No measurements or history in the source. (No source changes; the row keeps a size, as the table's other rows do.)
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines. (The diff names the public checkpoint and @Bizuayeu, as the rows already do.)

[ashhart]

Thanks @Bizuayeu.

The row and the recipe sentence match the card at 1c900ea0: 160.2 GiB of safetensors, and KL, top-1 and NLL all nearer BF16 than Jammed XSFP4, with paired intervals that exclude zero. "on main after 1.2.7 (#2894)" is the same wording Jammed XSFP4's row used before 1.2.3. Once #2894 ships in a release, the row and the recipe sentence can name that version, as #2451 did for Jammed XSFP4. The row stays as short as you meant it to, with the numbers left to the card.

CI is green on f048007945. It merges cleanly with main and with #2894's head 39abe29739. Since the row describes #2894's loader, it lands right after #2894 and not before. When you have your TP=2 serve on main (load time, memory, token hashes), please add it to the description. We won't hold the landing for it.

• Synthena (Ash's Agent)

[Bizuayeu]

When you have your TP=2 serve on main (load time, memory, token hashes), please add it to the description.

#2981 (comment)

Added to the Receipt, taken on main at 3885dceb2b with #2894's head merged, as #2894 is not in yet: loaded in 109.7s, rank 0's model 83.53 GiB, and the same token_sha256 drafted and with "draft": false on all three prompts. Once #2894 is in, I will serve it again on main.

[ashhart]

Thank you @Bizuayeu. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.