Skip to content
senn-techsenn-tech
Back to blog
AI Hardware
Local AI and GPU hardware2026-11-18· By Franz Senn

One day online, then pulled: NVIDIA's QAD checkpoint of Qwen3.8-Flash-Next

On Tuesday, 17 November 2026, NVIDIA listed an archive called nvidia/Qwen3.8-Flash-Next-NVFP4-QAD under its own org account in the Hugging Face Hub: a quantized variant of Qwen3.8-Flash-Next that NVIDIA refined further after quantizing it, of the model that has run on four RTX 5090 in our house since October. By Wednesday morning the archive was gone. We had read the file list and the safetensors headers at 06:15 UTC over HTTP Range, and around 06:30 UTC the same request answered with an error. Every NVIDIA figure here therefore carries one label: a one-time snapshot, not repeatable against the original.

A QAD checkpoint comes from a short training pass that runs after the quantization. A student model computes with fake-quantized weights during a few training steps, while a frozen teacher in BF16 supplies the targets. In the NVIDIA checkpoint that was 100 optimizer updates at a global batch of 512 sequences, as quantization_recipe.json described inside the archive. The stored weights were not quantized any further by that step.

Key points

  • The archive existed: two third-party repacks from 17.11., at 17:35 and 19:18 UTC, point via base_model at the NVIDIA revision a874339c; a trade article names 129.7 GB and 100 optimizer updates.
  • By 18.11. around 06:30 UTC the repo was gone: 401 on the page and in the API, absent from Hub search, never snapshotted by the Internet Archive. The September release still answers 200.
  • Until then we had read 129,780,932,631 bytes, that is 120.87 GiB across 36 shards with 300,217 tensors; routing experts and n-gram table are the same size at NVIDIA and in our build.
  • The licence file inside the NVIDIA repo read byte-identical to the Qwen upstream file, 3,235 bytes; today that is no longer checkable.
  • We never started the checkpoint, no card was free. Nothing changed in the production lane.

The sequence, as far as it is provable

Wordmark of Qwen
Qwen3.8-Flash-Next: 48 layers, 512 experts with 10 active per token, 262,144 tokens of context, plus a vision encoder and an n-gram embedding; our lane runs 220,000 tokens. (Quelle: Qwen / Alibaba Cloud, Wikimedia Commons, Apache 2.0)
  • 17.11., 09:32 UTC: creation date per the Hub API, read on 18.11. around 06:15 UTC. The day of its existence is independently proven by two third-party repacks from 17:35:51 and 19:18:26 UTC (Hub API), whose cards both name the full NVIDIA revision a874339cdb171ace5f36c494b6d0d3e54fdb92be as the base.
  • 17.11.: a trade article on explainx.ai describes roughly 129.7 GB of safetensors and 100 optimizer updates, noting that NVIDIA calls the accuracy gain statistically unproven.
  • 18.11., 06:15 UTC: our own measurement while the archive was open: the file list, all 36 safetensors headers over HTTP Range, plus hf_quant_config.json, quantization_recipe.json and the licence file.
  • 18.11., around 06:30 UTC: the same queries return 401, the model is missing from Hub search. Positive control right next to it: the September repo nvidia/Qwen3.8-Flash-Next-NVFP4 answers 200.
  • As of 18.11., 10:25 UTC: still 401, and never a snapshot in the Internet Archive, the availability API returns zero hits.

Whether NVIDIA deleted the archive or switched it to private cannot be told from the outside, and NVIDIA says nothing about it. The switch happened somewhere between 06:15 and 06:30 UTC, it is not measurable.

Size on disk: a small step since September

Checkpoint size on disk in GiB, our own measurement 18.11.2026Qwen original in BF16 (335.30)335.3NVIDIA NVFP4 from 2.9. (123.62)123.62Our build on 4x 5090 (122.54)122.54NVIDIA NVFP4-QAD 17.11. (120.87)120.870369
File sizes from the Hub file list and the referenced shards, our own measurement 18.11.2026. The QAD bar comes from 06:15 UTC, the repo has been unreachable since; the total size is corroborated by two third-party repacks and the explainx article. Distance from QAD to NVFP4: 2.3 percent, counted from the smaller build. (Quelle: Our own measurement Senn-IT, 18.11.2026)

Against the BF16 original that is 64.0 percent fewer bytes, against NVIDIA's first NVFP4 build from 2 September it is 2.75 GiB. QAD did not halve the size a second time: recipe and training pass only touched the precision of the smaller blocks, the blocks holding over 90 percent stayed as they were.

The component table, with its status

Componentour build (GiB)NVIDIA QAD (GiB)Difference
Routing experts63.2863.280.00
n-gram table (PLE)47.7547.750.00
Gated-DeltaNet3.892.01-1.88
Embedding and output head2.372.370.00
MTP module1.492.51+1.02
Other layers1.311.310.00
Attention (QSA)1.110.57-0.54
Vision encoder0.840.840.00
Shared experts0.440.12-0.32
QSA indexer0.040.02-0.02
Weight sum122.52120.78-1.74

Unit GiB (1024^3), measured from safetensors headers, measurement time 18.11.2026 around 06:15 UTC. The sum of 120.78 GiB is the pure weights; the file size of 120.87 GiB additionally holds 98,161,565 bytes of JSON headers. Ours sits at 131,579,038,666 bytes on disk, that is 122.54 GiB, with a row sum of 122.52 GiB. On weight level the two builds are 1.74 GiB apart, on file level 1.67 GiB, the difference sits in the shard headers. By data type our build holds 10.22 GiB in BF16, the QAD checkpoint 4.75 GiB, so 53.5 percent less; the rest moved into MXFP8 with block size 32. The QAD column is a snapshot from 18.11., confirmed by third sources only in the total size.

What the comparison lane reports, because it keeps running

All figures below sit in the start and runtime log of our machine. It launches via docker run in /home/bitnami/v032-test/run_port4.sh over the same safetensors index as the NVIDIA build, only local: 53 shards, 122.54 GiB of weights. The KV cache is capped at 7,600,000,000 bytes, that is 7.08 GiB and 1,351,242 tokens, 6.14 times the context. The n-gram table appears as weight_device=cpu, pinned=True, prefix hits around 80 percent, zero evictions.

VLLM_PLE_CPU_OFFLOAD=1 and VLLM_PLE_FORCE_FP8=1 sit in line 3 of the start script, plus one finding: VLLM_PLE_CPU_OFFLOAD appears in no file of the installed package in our production, the positive control hits VLLM_PLE_FORCE_FP8 and VLLM_KV_CACHE_LAYOUT, two of eleven files, and the boot log flags the variable with the warning "Unknown vLLM environment variable". The n-gram table still sits in pinned memory, decided by cpu_offload=True in engram.py. The switch is a dud, we did not remove it, it is an entry for the next maintenance round.

Recipe, release state and licence, as far as verifiable

The producer string in hf_quant_config.json read 0.46.0.dev281+g73d778422 when we looked, a development build. Against that, the release list of Model Optimizer, measured 18.11.: GA 0.46.0 from 18 August, 0.46.1 from 9 September, 0.47.0 from 23 September, 0.48.0 from 2 November, 0.47.1 from 6 November, and 0.49.0rc0 from 16 November as a prerelease. The dev build was cut before GA 0.46.0 and appears in no public release list; a rebuild from the public releases is not on the table, the training runs used an internal build.

The upstream card names the Qwen Community License, with a naming duty past 100 million monthly users and a special rule for renting the model as a service. The QAD card added an NVIDIA ownership reservation for its own changes, and its licence file was byte-identical to the upstream file, 3,235 bytes. The September build still carries no licence file, the API returns an entry error there; the adapter code in the vLLM repository runs separately under Apache 2.0. For external customer traffic, check this against your own contract.

What the incident leaves behind

After 18.11. only existence, date and total size remain provable, everything else stays a snapshot with a timestamp. Three questions NVIDIA leaves unanswered: the recipe ran on a dev build before the public GA, the card's effect claim was by its own account statistically unproven, and the licence wording of the experiment was identical to the September build. Why the archive vanished in under 24 hours remains open.

Why this story? Because it shows the bar we hold our own work to: numbers with a measurement carrier, a measurement time and a status, and a post that still stands when the source is unreachable. How we measure on this machine is the measurement series on vLLM 0.32 with MTP and NVFP4, why the n-gram table sits in our host memory explains the offload of the n-gram embedding. For vendor figures applies: vendor benchmarks against independent measurements. More from lane operations in the AI in operation hub, planning your own cards through the corporate LLM consulting.

Further Reading

Questions?
Why does Hugging Face no longer list the NVIDIA repo?+

We cannot answer that, NVIDIA says nothing about it. Measurable is this: the repo existed on 17.11.2026, and two third-party repos reference its revision a874339c. On 18.11. around 06:30 UTC our query got 401, while the September release of NVFP4 stays reachable. Deleting an archive and switching it to private cannot be told apart from outside.

What can still be checked without the repo?+

The total size of around 129.8 GB and the 100 optimizer updates, proven by two third-party repacks and a trade article from 17.11. The split by component, the recipe file and the byte equality of the licence are no longer checkable. Every figure about our own lane sits in the machine's start log and stays measurable there.

Does the incident change anything in our production lane?+

No. We never started either NVIDIA checkpoint, no card was free. The machine keeps running vLLM 0.32 on four RTX 5090 with MTP and NVFP4, the n-gram table in fixed host memory. The only open item is the maintenance entry about the PLE switch.