by Chris DePuy / September 22, 2026
Between September 9 and September 16, NVIDIA published first-party four-bit checkpoints of two frontier open-weight models: an NVFP4 build of Z.ai’s GLM-5.3-Flash and an NVFP4 build of DeepSeek’s DeepSeek-V4.1-Flash, both MIT-licensed, both produced with NVIDIA Model Optimizer, and both validated on rack-mount Blackwell silicon rather than on the desk. That distinction is the story. The vendor’s own validation envelope names GB200 and GB300 hardware; the receipts showing the same checkpoints serving on a four-node DGX Spark fleet, with per-step timings and failure modes, are entirely community-published. We have tracked the deskside quantization beat since the first community NVFP4 experiments, and this is the first round in which the vendor’s shipped weights and the community’s served weights can be read against one another.
NVIDIA: what the conversion actually changes
The DeepSeek checkpoint converts only the ordinary routed experts — projections w1, w2 and w3 across 384 experts in 40 layers — from the source MXFP4 format to NVFP4 weights and activations at group size 16, per the model card; attention, shared experts, the vision encoder, the Engram lookup tables and the multi-token-prediction and DSpark drafter tensors retain their source precision. The card records all 16,986,931,200 weight blocks as passing lossless conversion, with 18 of 46,080 projection entries falling back to default scales, and it states the consequence plainly: the finer NVFP4 scale layout grows the export from approximately 476 GiB to 492 GiB across 48 safetensors shards. A larger file, in other words — the vendor traded footprint for Blackwell-native execution, not for bytes.
The GLM package quantizes a different slice. Its card covers the linear operators inside the sparse-MoE shared experts and the dense MLPs of the 320-billion-parameter, 18-billion-active model, cutting bits per parameter from 16 to 4 and disk and GPU memory demand by roughly 3.33x under the recipe nvfp4_experts_dense_mlp-kv_fp8_cast. Both packs were calibrated on the CNN/Daily Mail corpus and NVIDIA’s Nemotron post-training set; for the DeepSeek export, the card specifies 512 samples per dataset and a sequence length of 512.
Vendor evaluation: parity, with one exception each
On the published tables the conversions hold their accuracy. For DeepSeek-V4.1-Flash, the card lists GPQA Diamond at 91.035 rising to 91.288, AA-LCR at 78.563 slipping to 78.438, SciCode 54.401 to 55.843, IFBench 76.667 to 77.267, MMMU-Pro 74.046 to 73.699 and Terminal-Bench 2.1 81.60 to 82.16, all measured with vLLM on four GB300 GPUs. For GLM-5.3-Flash, tested on GB200, the paired readings run GPQA Diamond 0.9217 to 0.9211, SciCode 0.5621 to 0.5769, MMMU-Pro 0.7688 to 0.7630, AA-LCR 0.71 to 0.7106, IFBench 0.613 to 0.6054 and Terminal-Bench 2.1 0.8258 to 0.8315. Six benchmarks, two models, no collapse — and the DeepSeek card volunteers the caveat that “lossless weight conversion does not imply identical inference outputs,” which is the most honest line in either release. The wider architectural context is in DeepSeek’s own paper, submitted September 17: an 890-byte-per-token global KV cache, about one quarter of the previous generation’s, with persistent cache on disk cut to roughly one eighth via SWA bounded replay.
The community: four DGX Sparks, two receipts, one bug each
The deskside ports arrived fast and, in one case, expensively. The four-Spark DeepSeek recipe documents runs/2026-09-17-nvfp4: fourteen launches of the official checkpoint, none of which reached a serving state; the maintainers isolated the memory mechanism — private mappings of the shards broken under copy-on-write during the fused-MoE expert copy — and a single lever that held, expert parallelism with the weight filter enabled, skipping 194.6 of 302 GiB per rank. Read that against the same lane’s served results the following week: cold prefill at 1,939 to 2,048 tokens per second, time to first token 0.205 s at one stream, and 475 tokens per second aggregate for coding at twelve concurrent streams.
The GLM recipe now ships the official pack as its default lane, with the maintainers’ own NVFP4 quantization layered on the 403 attention tensors that the upstream export leaves in BF16 — two patched files stop the engine from forcing the attention projections’ quantization config to null, because a packed four-bit weight otherwise has nowhere to load and the stack fails to boot entirely. Measured on one fleet at a 500,000-token window with an FP8 pool of 3,532,196 tokens at 43.76 GiB per rank: cold prefill 1,997 tokens per second on a 40,659-token prompt, and seven of nine single-stream categories within measurement noise of the previous community checkpoint (the run log flags those two single-category readings as landing outside their own spread bands, one high and one low), with concurrency ahead at every level tested (up 4.8% at twelve streams, 6.9% at sixteen, 7.6% at twenty-four, 10.0% at thirty-two) and the drafter’s acceptance measured at 0.394 with the DFlash2 block-diffusion proposal at depth seven, per the operator’s serving post. A de-risked uncensored derivative, Blackfrost’s NVFP4 card, measures at or above the shipped lane on almost every measure, mathematics the single category left behind (83.3 against 88.8, at the edge of its spread), at a draft acceptance of 0.415 against 0.394 — and it arrives with a bug of its own: per the boot log, its config declares the four-bit-activation path while shipping zero input-scale tensors, a mislabeled string the recipe rewrites to weight-only intent before the stack will boot.
| Lane | Nodes | Cold prefill | Decode evidence | Failure note | Source |
| NVIDIA validation, DeepSeek pack | 4x GB300 | accuracy tables only | vLLM, SGLang smoke tests pass | not served on GB10 in vendor tests | model card |
| NVIDIA validation, GLM pack | 4x GB200 | accuracy tables only | vLLM, SGLang | not served on GB10 in vendor tests | model card |
| Community port, DeepSeek pack | 4x DGX Spark | 1,939–2,048 tok/s | 475 tok/s aggregate, 12 streams | 14 boots failed before one lever held | recipe |
| Community port, GLM pack | 4x DGX Spark | 1,997 tok/s | +4.8% to +10.0% concurrency vs prior | 403 tensors quantized locally, 2 files patched | recipe |
The analytical reading: the official checkpoint is now the default, and the community’s value-add has migrated from the quant itself to the serving apparatus around it. Vendors validate on the datacenter part of the market; the deskside evidence — fourteen failed boots, one held lever, 403 hand-quantized tensors — is what makes a 492 GiB package servable on 128 GiB of unified memory at a desk. Independent operator reports keep the picture moving: one operator post on the same DeepSeek build reports four Sparks serving while throughput work continues, and another operator’s uncensored lane outperformed the censored one on nearly every prompt category. Two findings that should keep the procurement question open: what to serve is increasingly decided by who patched the loader, not by who exported the weights.
As firms such as 650 Group have been tracking in their AI infrastructure research, the first-party-versus-community boundary in local inference is dissolving one checkpoint at a time — the vendor now ships what the community once reverse-engineered, and the community’s remaining moat is the boot log.
A separate batch of .md notes on the open decision-model round — a 0.6B/4B/8B family of Jev-like classifier models with out-of-domain scores of 79.6% against 85.7% for the hosted original, a Claude-Code-side clone of the same category, and a diffusion-language vLLM patch reporting some 120 ms single-request decision latency — is pending a dedicated post, as is the new single-GB10 engine posting some 1,126 tokens-per-second prefill at an 8K prompt and a 141,519-token conversation restored from disk in 2.1 seconds, and the single-Spark Qwen3.8-Flash-Next round now serving 117 tokens per second of prose and 180 of code at eight streams with peak memory down from 101 to 91 GiB.