Qwen3.8-Flash-Next: 6 billion active, 180 to load, 51 sitting in a table

On 24 August 2026, at 08:24:59 UTC, the Qwen organisation — Alibaba’s open-weight brand — published Qwen3.8-Flash-Next on Hugging Face: not the next version of the Qwen3.8 family, but, as the card itself states, “this experimental preview of the architecture that will underpin Qwen4”. For anyone assessing an open release to run inside their own perimeter, a preview declared as such changes the question: not only how much it weighs and what licence it carries, but whether an architecture the vendor itself calls experimental belongs today in an environment that matters. We downloaded the card, the configuration file, and the header of each of the 131 weight files. We verify it here, as of 30 August.

The facts, in order

Publication: repository created on 24 August at 08:24:59 UTC, last updated on 27 August at 05:03. Organisation Qwen: verification badge absent (isVerified: false), team plan, 464 published models, 101,306 followers.
Access: gated: false. Verified without authenticating: the card (HTTP 200) and the header of a weight shard (HTTP 206, partial response) are both reachable with no agreement to accept.
Adoption, today: 52,341 downloads, 4,285 likes, read on 30 August — wide adoption for a six-day-old release.
Architecture: Qwen4ExpForConditionalGeneration, type qwen4_exp — the card calls it “the first open-weight release under this architecture”.
Weight on disk: 131 .safetensors files, 359,999,963,128 bytes in total per model.safetensors.index.json — 360.0 GB, 335.3 GiB. Verified by reading the header of each of the 131 shards (not the full weights): every tensor is BF16, bar 280 bytes stored as int64.

Six active, 180 to load: where the other 174 sit

For the language model alone, the card states: “Number of Parameters: 125B with 6B activated, plus 51B n-gram embedding and 4B MTP”. Summing the declared bytes across the 1,658 tensors in the 131 files, grouped by name, our own count gives: 125,743,653,795 parameters in the model body (125B declared), 51,200,245,760 in the n-gram table (51B declared — a near-perfect match), 2,607,150,848 in the MTP module (4B declared — here the sum does not add up). Added together, the three language-model figures come to 179.55 billion, against the 180 implied by the declared sum (125+51+4): the gap sits almost entirely in the MTP module. The vision encoder adds a further 448,931,056 parameters, not covered by that figure: the total across everything in the 131 files is 180.0 billion. We do not know whether the card, for the MTP module, counts shared weights that the tensor names do not separate out.

The n-gram table is the piece that matters for anyone sizing hardware. The configuration file places it at the third layer (ple_layer_ids: [2], index 1 in the tensor names) with a vocabulary of 20,000,000 bigrams and trigrams (ngram_vocab_size_base) — exactly the figure the card states. It is not a compute block: it is a lookup table, indexed by n-gram rather than by token. The advantage, the card writes, is precisely that: “Embeddings provide a unique axis for parameter scaling that requires less computation and is more amenable to offloading than Mixture-of-Experts (MoE)”. Less compute does not mean less memory that has to be kept ready: 51.2 billion parameters, 102.4 GB in BF16, must be loaded regardless, even though the model activates only 6 billion per token. The active-to-total ratio on the language model alone is 1 to 30 — more extreme than what we measured this month on another open-weight model, where the ratio was about 1 to 12, albeit at a much smaller overall size.

The hardware sum

360 GB in BF16 does not fit on a single 80 GB GPU, nor on a 141 GB one: it takes several cards, or a quantised version — the compute that makes downloaded weights runnable is a separate sum, and skipping it does not buy you sovereignty. Qwen published one twenty-four seconds later, Qwen3.8-Flash-Next-FP8: 185,502,232,570 bytes, 185.5 GB, 172.8 GiB — nearly half, verified on the same index file, but still beyond a single card’s capacity if you want headroom for a long context. The native context, 262,144 tokens per max_position_embeddings, matches what the card states; the declared extension to 1,000,000 is not an ambiguity of the kind seen in other recent cases: Qwen explains how to get there, with a YaRN factor to set by hand in rope_parameters, and warns that performance on shorter texts suffers as a result. The repository ships no dedicated modeling_*.py file: whoever runs it depends on an engine that already recognises qwen4_exp, a model type that is only days old. The configuration file records the library it was saved with, transformers==5.8.0.dev0 — a development build from that moment, not necessarily a requirement today for anyone using the three engines the card itself recommends for production: SGLang, vLLM, TokenSpeed.

The licence, briefly

Not Apache 2.0: license: other, license_name: qwen-community-1.0, the full text downloaded (15 lines, HTTP 200). Two conditions: above 100 million monthly active users or US$20 million in monthly revenue, the model name must be shown in the interface; and — here without the revenue threshold we verified on the flagship model of the same family — anyone running a Model as a Service or AI Work Assistant business must first obtain a separate licence from Qwen: “The foregoing requirement shall not apply to the licensee’s internal Use of the Software, provided that such Use does not make the Software, its outputs, or its underlying model capabilities available to any third party.” An organisation that today uses the model purely in-house stays outside the clause; the day that service is shared with another department or offered to a client, the threshold applies without any revenue figure needing to be crossed.

What we do not know

We have not run the model: no claim of ours about quality or real-world behaviour. The benchmark comparisons against Qwen3.8-27B, Qwen3.7-Plus, DeepSeek-V4-Flash and Claude Opus 4.6 are the vendor’s own tables, not an independent trial. We do not know how Qwen arrived at the declared 4B for the MTP module, nor whether it attributes shared weights that the tensor names do not separate. The card does not specify the provenance of the training data. We offer no legal qualification of the licence. And we do not know when, or whether, today’s experimental architecture becomes the stable one behind Qwen4: the card itself makes no such promise.

How we solve it

A preview declared as such is neither dismissed nor adopted out of habit: it enters the model registry with the measured figures — not the ones rounded by the announcement — version, digest, real footprint per active token and per byte on disk, verified inference engine, the licence text archived exactly as it read today. That is the trail an inspection needs, and the basis on which the same set-up holds models, data, archives and decisions together in a single operating model, where AI agents execute with an operator in command: for large enterprises, defence and public administration. On-premise on self-contained machines that require no deep integration into the client’s network, or a dedicated cloud with a data centre in Italy — always with shared management, because a days-old architecture is not the moment to improvise who administers it.

See the service · Talk to an engineer

Is someone in your organisation already assessing an architectural preview downloaded from Hugging Face for a use that matters? Half an hour with one of our specialists, at no cost: we bring the real sum on memory, licence and inference-engine maturity before it reaches production.