[Bizuayeu]

What this changes

Two rows in the README's Pick a model, after GLM-5.3-Flash NVFP4, and one sentence in the GLM CUDA recipe's opening. Main loads NVIDIA's FP4-Local checkpoint as stored since #1050 (b34f9e781..e2f1e2cd9), and XSFP4, a derivative of FP4-Local that keeps its tensors outside the routed experts, loads through the same path. Both rows say "on main after 1.1.4" until a release carries #1050.

The rows give each checkpoint's size and how close it comes to BF16. FP4-Local, NVIDIA's mixed-precision checkpoint, is 13% smaller than NVFP4 and as close to BF16. XSFP4 is FP4-Local with its routed experts' NVFP4 blocks fitted again from the BF16 weights: it is within 0.4 GiB of FP4-Local's size and closer to BF16 than both at TP=2.

The pulls pin FP4-Local at 76348ddb, the revision #1050's receipt read, and XSFP4 at 88736f91, the revision the runs below read. XSFP4's later commit b761a3a5 changes its model card alone.

Receipt

Documentation only, which needs no receipt. The rows quote these:

• Sizes: the safetensors files at the pinned revisions on Hugging Face. NVFP4 (423acf37) has 190.40 GiB, FP4-Local (76348ddb) 165.28 GiB and XSFP4 (88736f91) 165.69 GiB.
• Closeness to BF16, next-token distributions over the same 128 documents (144,285 positions):
  • glm/cuda: load and serve nvidia/GLM-5.3-Flash-FP4-Local as stored #1050's receipt (s5/): NLL above BF16 of +1.11% for FP4-Local and +1.37% for NVFP4.
  • XSFP4's card measured all three the same way on 1.1.2 with glm/cuda: load and serve nvidia/GLM-5.3-Flash-FP4-Local as stored #1050's code. At TP=2, top-1 agreement, NLL above BF16 and the KL lower bound are 93.26%, +0.85% and 0.0431 for XSFP4, 92.82%, +1.11% and 0.0510 for FP4-Local, and 92.68%, +1.37% and 0.0510 for NVFP4. The per-document 95% intervals put XSFP4 closer than both by all three measures, and do not tell FP4-Local and NVFP4 apart by the NLL or the KL bound. At TP=3, XSFP4 is closer than both by top-1 agreement and the KL bound, and closer than NVFP4 by the NLL.
• FP4-Local on the engine: #1050's receipt on the pull request's head 80517fc3, at TP=2 and TP=3. It was not rerun on main.
• XSFP4 on main 8086c205c, with the engine built from it for sm_121 (Zig 0.17.0, CUDA 13.3.1). One start at each size, with drafts on, --parallel 1, --vision and TF_KV_PAGES unset:
  • TP=2, two GB10s over a direct link, --context 524288: loaded in 370.7 s. Rank 0 reported model 96.97 GiB; 1 stream at once, 6.71 GiB each at a 524288-token window, of 5.3 GiB left after a 12.2 GiB reserve and a 5.3 GiB prompt cache.
  • TP=3, three GB10s over a ring of direct links with /32 addresses, on glm cuda: rank 0's NCCL id goes over the ranks' lines, so a /32 ring's followers bootstrap on NCCL_SOCKET_IFNAME #1178's bootstrap, --context 1048576: loaded in 313.8 s. Rank 0 reported model 63.60 GiB; 1 stream at once, the engine's own and 0 more taking 2049 pages of 512 rows (12.76 GiB) as they need them (enough for one at a 1048576-token window).
  • At both sizes, a greedy /v1/completions of 16 tokens returned the same token_sha (e49d850e547e).
• The new rows' relative links resolve in this tree, the recipe's lines stay within 120 columns, and git diff --check is clean. Against each open pull request that touches README.md or the recipe (flint: serve Recursive Flint (40B-A2.5B MoE) natively on Metal #1177, qwen27: TF_Q27_PROFILE logs the round's class GPU times, and replies carry gpu_seconds #1215, minimax_m3: serve MiniMax-M3.1-Flash (428B-A23B MoE, images) natively on Metal #1235, drama: add native Metal NVFP4 and reference CFG support #1268, cosmos3: Cosmos3-Edge reasoner (image + text) on Metal (new family, proposal for #1269) #1271, cosmos3: Cosmos3-Edge reasoner on CUDA (stacked on #1271, proposal for #1269) #1272, server: expose per-request Drama speech sampling controls #1277, server: accept dialogue speaker tags (depends on #1277) #1278), git merge-tree reports the same conflicting files on this branch as on main.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off. (Documentation only; no code changes.)
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (No code changes. The XSFP4 runs above were a start and one greedy completion at each size, not an exactness check; FP4-Local's checks are in glm/cuda: load and serve nvidia/GLM-5.3-Flash-FP4-Local as stored #1050's receipt.)
• No precision traded for speed. If the bits change, the description says which sums change and why. (No code changes; no bits change.)
• A new or changed GPU kernel is registered with the kernel contract (docs/kernel-contract.md); each backend's SUMMARY line is in the receipt. (No kernel added or changed.)
• Prompt processing is no slower than the last release. (No code changes.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not. (CUDA on GB10 only, at TP=2 and TP=3, for XSFP4. No Metal and no other GPU.)
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No tests added; documentation only.)
• Comments and docstrings are one line. No measurements or history in the source. (No source file touched.)
• No personal data, machine names, internal hosts or local paths. No AI attribution lines. (None in the rows, the recipe or this description.)

[ashhart]

Landed on main as 8286ce1 (f3c1ab0..67cc8fb). Thanks for the FP4-Local and XSFP4 rows and the measured loads at TP=2 and TP=3.

- Synthema, Ash's AI Assistant.

[Bizuayeu]

Landed on main as 8286ce1 (f3c1ab0..67cc8fb). Thanks for the FP4-Local and XSFP4 rows and the measured loads at TP=2 and TP=3.

#1285 (comment)

Thank you for landing it, and for shipping it in 1.1.5. Seeing XSFP4 in Pick a model, loaded natively by a release, means a lot to me: it is a real joy to see something I made show up in TensorFold as a model anyone can pick. It also feels like the start of a time when a model's weights and the engine that serves them are shaped together.