ZwZ2-9B vs Qwen3.8-27B: The Small Slot the New Qwen Generation Left Open

The Qwen3.8 generation shipped with a 27B dense model on 5 August 2026, a 2.4-trillion-parameter mixture-of-experts on 8 August, and a Flash-Next variant on 24 August. It did not include a 9B. That means the small end of the current line-up is still occupied by Qwen3.5-9B, a checkpoint from 27 February 2026 — and it means ZwZ2-9B, posted by inclusionAI on 16 October 2026, sits on the newest available Qwen3.5 stack at that size. Compare it against Qwen3.8-27B and you are not comparing two models from one family. You are comparing a gap of roughly eighteen billion parameters inside one architecture, because the 3.8 generation never renamed it.

The thing to understand first is that this is not a rivalry in any ordinary sense. Qwen3.8-27B is a finished product with a licence, a hosted API, published evaluations and seven million downloads in the last month. ZwZ2-9B is an undocumented derivative checkpoints in a repository with no README. The useful question is narrower and more practical: if a 27B model is the one Qwen wants you to run, what exactly are you buying by staying at 9B, and what does the cheaper checkpoint inherit from the generation above it?

The architecture never changed — only the scale did

This is the finding that makes the comparison worth writing, and it is visible in two config files. Qwen3.8-27B declares its architecture identifier as qwen3_5. So does ZwZ2-9B. Same model type, same class name, same attention design, same tokenizer lineage — a generation apart, and the label never moved.

Look past the identifier and the family resemblance is complete. Both use the same three-to-one schedule of linear-attention layers to full-attention layers. Both reserve a 248,320-token vocabulary and declare a 262,144-position context ceiling. Both use interleaved M-RoPE with the same section split and the same 10,000,000 rotary base. Both carry a vision tower of exactly 27 layers at hidden size 1,152 with 2,304 position embeddings. Qwen3.8 is not a new design that superseded Qwen3.5. It is the same design at larger dimensions, and Qwen3.8-27B says so in its own config.

• Parameters — ZwZ2-9B: 9,409,813,744. Qwen3.8-27B: 28 billion per its card, whose config specifies 64 layers at hidden size 5,120 against 32 layers at 4,096.

• Architecture identifier — qwen3_5 in both. Same class, same attention schedule.

• Vocabulary — 248,320 in both, image and video token ids included.

• Context ceiling — 262,144 positions declared in both, with Qwen's hosted version extending to 1M.

• Vision tower — 27 layers at hidden size 1,152 in both; the projection widens from 4,096 to 5,120 to match the language stack.

• Reasoning default — ZwZ2-9B answers without thinking unless asked. Qwen3.8-27B thinks by default, with a tunable reasoning effort.

• Licence — none declared vs Apache-2.0.

• Evidence — a config file vs a model card, evaluation results, a hosted API and 198 published adapters.

What the 27B buys you that the 9B derivative does not

The first thing is permission. Qwen3.8-27B is Apache-2.0 with an explicit grant, deployable on Azure and SageMaker according to its own tags, and served through Qwen's official hosted API with a 1M-token context in the managed version. ZwZ2-9B declares no licence at all, which is a different problem from a restrictive licence — it is the absence of a grant rather than a narrow one, and it is the single largest risk in adopting it.

The second is reasoning control. Qwen3.8-27B's chat template exposes three reasoning-effort levels with a default of the highest, instructs the model accordingly, and supports preserving reasoning context across turns. It also keeps the Qwen convention of thinking by default. ZwZ2-9B has none of this machinery, and its template inverts the default — it emits an empty <think></think> block unless you explicitly pass a flag, which means the model answers directly and cheaply, and gives you no dial to turn when a question deserves more thought.

The third is the ecosystem, which is the least discussed and most load-bearing. The Hub lists 198 adapters, 553 fine-tunes, 24 merges and 1,472 quantisations built on Qwen3.8-27B, plus a hundred Spaces. That is a body of work by people who are not Qwen, describing how the model behaves on real problems. ZwZ2-9B has none of it, and cannot have any of it yet — it is hours old.

What staying at 9B still gets you

There is a reason the small slot has not gone quiet. Qwen3.5-9B, the checkpoint ZwZ2-9B is built on, recorded 8,591,266 downloads in the last month — more than the 27B flagship's 7,363,968 over the same period. The 9B size is where a great deal of real work happens: it is the largest model that fits comfortably on a single consumer accelerator, it is the size teams fine-tune when they want their own weights rather than an API, and it is where an undocumented derivative like ZwZ2-9B can plausibly be interesting, because the thing it inherits has been read by millions of people.

Stripping the multi-token-prediction head is also not obviously a loss. ZwZ2-9B's config sets mtp_num_hidden_layers to 0 where Qwen3.5-9B has 1, which removes a speculative-decoding head and roughly a quarter of a billion parameters with it. Whatever the training run intended, the artefact you download is a simpler model than its parent, and simpler is easier to reason about when you cannot read a card.

And the vision path is the same in both directions: 27 layers at hidden size 1,152, a 2,304-position embedding table, a video processor taking up to 768 frames at 2 frames per second. If you need images in, the 9B does not give that up to the 27B.

What this looks like if you are choosing an endpoint rather than a checkpoint

The practical answer for most teams is not either-or, and the hosted side of it is settled in a way the open-weights side is not. Qwen3.8-27B is routable on OrcaRouter as qwen/qwen3.8-27b — a 262K-token context, text, image and video in, text out, at $0.33 per million input tokens and $2.40 per million output. That is the provider's list price passed through with 0% markup, so when Qwen changes a price the number here changes the same day rather than at the next catalogue refresh, and it sits next to qwen/qwen3.8-flash and qwen/qwen3.8-max on the same key if you want to compare tiers on your own traffic rather than on someone's benchmark table.

ZwZ2-9B is not in that catalogue and we are not going to imply otherwise — it is a Hugging Face download with no serving endpoint published by its vendor, and inclusionAI distributes through the vendor's own channels. If you want a hosted stand-in for the small multimodal slot while you evaluate the checkpoint, qwen/qwen3-vl-8b-instruct is an 8B vision-language model with a 131K-token context at $0.18 and $0.70 per million tokens, documented and licensed. Put a candidate behind the same interface as the incumbent and comparing the two costs a model string; that is the entire argument for routing a model you are not sure about instead of integrating it.

The generational question, answered honestly

There is no benchmark on which ZwZ2-9B beats Qwen3.8-27B, because there is no benchmark on which ZwZ2-9B has been measured at all. A 28B model with a technical card and a 9.4B checkpoint with a config file do not compete; one of them has been evaluated and one of them has been described. Anyone producing a head-to-head result this week is describing their own impression of a model they downloaded this morning.

What is genuinely new here is the lineage, and it is worth stating plainly because it is easy to miss. The Qwen3.8 generation did not replace its architecture; it scaled it. Qwen3.8-27B and ZwZ2-9B answer to the same architecture identifier, share a vocabulary, share a context ceiling, share an attention schedule, and share a vision tower down to the layer count and the hidden width. That is useful knowledge in both directions: it tells you what a 9B checkpoint on this stack is likely to feel like, and it tells you that the gap between the small and large ends of the Qwen family is a matter of degree rather than a change of species.

So the honest verdict on ZwZ2-9B versus Qwen3.8-27B is that the 27B is the one you can plan around, and the 9B derivative is the one worth watching. Read the config, run your own task against it, and wait for a licence before it touches anything that matters.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.