ZwZ2-27B vs Qwen3.8-27B: Same Skeleton, Two Very Different Bets

The strange thing about ZwZ2-27B versus Qwen3.8-27B is that the two models are the same machine. Open both config files and the numbers line up line for line: 64 layers, 5,120 hidden size, 24 query heads against 4 KV heads at dimension 256, a 17,408 feed-forward width, a 248,320-token vocabulary, 262,144 positions, and a 27-layer vision tower at 1,152 hidden size. They even declare the same architecture class. What separates them is not width or depth. It is what each lab did to that body afterwards — and, so far, nobody has run them against each other to find out which choice won.

That absence is the first thing to say plainly. inclusionAI's published evaluation for ZwZ2-27B compares it with GPT-5.6-Luna, GLM-5.3-Flash and DeepSeek-V4.1-Flash. It does not include Qwen3.8-27B, and it does not include any Qwen model newer than the one ZwZ2 was built from. Qwen, for its part, has never mentioned ZwZ2 at all. Every number below is one lab measuring itself, or Qwen measuring itself, and no third party has put the two in one harness.

Same skeleton, line by line

It is worth being precise about how close these are, because the usual “27B versus 27B” framing understates it. These are not two models that happen to land at similar scale; they are two post-training runs over an interchangeable architecture.

• Parameters — 27B for both, dense, no mixture-of-experts routing in either config.

• Depth and width — 64 layers, 5,120 hidden size, 17,408 intermediate size, identical in both files.

• Attention — 24 query heads, 4 KV heads, head dimension 256, partial rotary factor 0.25, rotary base 10,000,000.

• Layer schedule — a 3:1 linear-to-full attention pattern, 16 repetitions, in both.

• Vocabulary and context — 248,320 tokens and 262,144 native positions, in both.

• Vision tower — 27 layers, 1,152 hidden size, projecting to 5,120, in both.

The practical consequence is that the hardware question is the same question for both. Roughly 54 GB of weights in bf16, one 80 GB card comfortably or two 40 GB cards tightly. Neither is a model you run on a laptop, and neither has a quantized release in its own repository — Qwen's ecosystem produces those, and ZwZ2 has almost no ecosystem at all.

The one line that differs, and two defaults that matter more

Structurally there is exactly one divergence: Qwen3.8-27B sets mtp_num_hidden_layers to 1, keeping a multi-token prediction head that Qwen says was trained with multiple steps. ZwZ2-27B sets it to 0. The inclusionAI model has had the speculative-decoding head amputated, which is what you do when your training pipeline does not use it and you would rather ship a smaller, simpler artifact. It is a housekeeping difference, not a design statement.

The defaults are where the two models actually part company, and they will shape your integration more than any config value.

Thinking. Qwen3.8-27B thinks before answering by default, exposes three effort levels with xhigh as the default, and can retain reasoning context from earlier turns via preserve_thinking. ZwZ2-27B is documented in non-thinking mode: its quickstart passes enable_thinking=False, its chat template asks the model to place final answers inside <answer> tags, and the entire 25-benchmark evaluation was run that way. If your pipeline expects a single-pass response, ZwZ2-27B is shaped for you out of the box. If you want the model to deliberate across a fifty-step task, Qwen3.8-27B has machinery for that and ZwZ2-27B's card does not claim any.

Modality. Qwen3.8-27B is documented for image and video understanding, with video sampling controlled through the serving layer. ZwZ2-27B's config carries a video_token_id, but its card and evaluation are entirely about still images and region comparison. Treat video as unproven on the inclusionAI side.

What each side can actually prove

Both vendors published evaluations. They are not comparable, and the reason is not marketing — it is that the two suites measure different things.

ZwZ2-27B's claim lives on 25 benchmarks in six perceptual domains, scored by inclusionAI, moving its own base from 55.52 to 60.51. The strongest absolute numbers are in fine-grained perception (67.57 domain average) and long-context understanding (68.77), and the weakest are Chartography at 12.00 and AgentVista at 18.18.

Qwen3.8-27B's published table is a text-and-agent suite with a vision section, and Qwen ran it against its own predecessor. Some of the gaps between Qwen3.8-27B and Qwen3.6-27B are enormous:

• Agentic terminal coding — Terminal Bench 2.1 at 73.0, against 63.4 for Qwen3.6-27B.

• Software engineering — QwenSWEBench at 79.0 against 49.3, and DeepSWE 1.1 at 42.2 against 13.3.

• Computer use — OSWorld-Verified at 84.3 against 63.9; WebArena-Verified at 64.8 against 48.8.

• Mobile use — AndroidWorld at 81.9 against 70.3.

• Multimodal software engineering — SWE-MM at 38.6; RecreationBench at 47.1 against 29.8.

• Scientific reasoning — GPQA Diamond at 89.2 against 87.8, and HLE at 30.8 against 24.0.

Those are Qwen's own numbers on Qwen's own harness, and they say something simple: the 3.8 generation is not a small step over 3.6. On agentic software work the improvement is measured in tens of points, not single digits.

Why +4.99 against 3.6 is not a number about 3.8

This is the trap in the comparison, and it is easy to fall into. ZwZ2-27B's headline gain is +4.99 over Qwen3.6-27B. Qwen3.8-27B is a much stronger model than Qwen3.6-27B on Qwen's suite. The tempting inference — that ZwZ2's gains close the gap — does not follow, and there is a mechanical reason it cannot be checked.

You cannot subtract across runs. ZwZ2-27B's 60.51 is an average of accuracies over 25 perceptual benchmarks chosen by inclusionAI; Qwen3.8-27B's figures are scores on a different set of benchmarks chosen by Qwen. A model scoring 60.51 on one suite and a model scoring 84.3 on a completely different task tell you nothing about their relative order, and any article that puts those numbers in one column is manufacturing a ranking that no one measured.

What can be said is narrower and still useful. ZwZ2-27B is a targeted intervention on a model Qwen has since superseded, and the intervention's documented size — five points on a perceptual mean — is the same order of magnitude as several of Qwen's own 3.6-to-3.8 deltas on vision tasks, and far smaller than its deltas on agentic tasks. Whether the specialization clears the generational gap on region-level perception is exactly the thing that would need a shared harness. Nobody has built one.

The card's reference columns hint at why an answer might exist. ZwZ2-27B scores 60.51 where inclusionAI measured GLM-5.3-Flash at 57.24 and DeepSeek-V4.1-Flash at 55.47 — hosted models it beat on a suite built around region comparison, with the card noting that advanced models' image resolution was constrained by their serving platforms. That is a result about the suite as much as about the models. It is not evidence about Qwen3.8-27B, which was never in the room.

The license decides it for most teams before the benchmarks do

Qwen3.8-27B is Apache-2.0, declared in the model card's front matter, with a downloads count in the millions and an ecosystem of adapters, fine-tunes and quantizations around it. ZwZ2-27B declares no license at all. Its card points to “the license accompanying the released weights and the applicable backbone license,” which is circular when the weights carry none.

For a research team that is a footnote. For a company putting a model behind a product it is the deciding line, and it is not close. The backbone being Apache-2.0 does not automatically license a derivative whose own terms are unstated — that is a question for counsel, and until inclusionAI answers it, the safe reading is that ZwZ2-27B is not cleared for commercial deployment. The absence of a paper compounds it: the card still contains the literal placeholder PAPER_URL, so the method description is a README and a dataset schema, nothing more.

What you can call today

Qwen3.8-27B is routable on OrcaRouter right now, under one key with the rest of the catalogue and at provider list price with no markup — so if Qwen changes its pricing, the change is live here the same day rather than after a repricing cycle. It is listed at $0.33 per million input tokens and $2.40 per million output tokens, with 262,144 tokens of context on offer. That matters for a comparison like this one, where the cheaper experimenter usually wins: you can benchmark Qwen3.8-27B against your own workload this afternoon without a download or a GPU.

ZwZ2-27B is not in our catalogue, and it would be dishonest to phrase that more softly. It is a 54 GB self-hosted artifact with no license and no serving recipe, and the only way to evaluate it is to run it. If it earns a place in your stack later, the routing layer is where the two would meet — one endpoint, automatic failover, and a policy that sends the region-comparison traffic to the specialized model and everything else to the generalist, rather than a rewrite.

Which one, by workload

Qwen3.8-27B is the defensible default for almost everything. Agentic coding, long-horizon tool use, computer and browser control, general reasoning, video — that is where its published gains are largest, and it is the only one of the two you can legally ship and practically deploy today.

ZwZ2-27B has a narrower case: you already run a Qwen3.6-27B-class pipeline, your bottleneck is measured detail in documents, charts or interface screenshots, and the idea of adding zoom-and-crop tool calls to your inference loop is unappealing. That is precisely the problem inclusionAI claims to solve — learn the comparison across marked regions at training time so that inference needs no tools. If the base model is the control condition in your own harness, a local test of ZwZ2-27B against it is a cheap and well-posed experiment.

What is not defensible is treating the vendor's 60.51 as a general capability score, or comparing it with any Qwen3.8 figure as though the two sat on one scale. They do not, and the model that would settle it — a shared harness, run by someone with no stake — has not been announced any more than ZwZ2-27B itself was.

Three questions this comparison really raises

Is ZwZ2-27B simply Qwen3.6-27B with extra training? Architecturally yes, down to the layer schedule and vision tower, with the multi-token prediction head removed. In behaviour it depends on a post-training recipe inclusionAI has described only at the level of a README: region groups of three to seven, questions generated from privileged crops, filtered by cross-model agreement and drop-one-region tests, then GRPO on about 20,000 examples with no supervised fine-tuning stage. That is a description of a method, not a reproduction.

Would ZwZ2-27B's gains survive contact with Qwen3.8-27B on perception tasks? Unknown, and unknowable from the current record. Qwen3.8-27B's vision section is weighted toward computer use, browser use and multimodal software engineering, not toward region-level comparison, so even Qwen's own table does not answer the question the ZwZ2 claim raises.

Does the thinking-mode default change the comparison? Yes, more than the benchmark gap suggests. ZwZ2-27B is evaluated and documented as a single-pass, non-thinking model; Qwen3.8-27B is built around adjustable reasoning effort. For latency-sensitive perception calls that favours ZwZ2-27B's design, and for anything requiring deliberation it favours Qwen3.8-27B's — but the two were measured under different generation settings, so the benchmark numbers carry that difference too.