ZwZ2-27B vs ZwZ2-9B: What Three Times the Parameters Actually Buys

On exactly one benchmark in inclusionAI's 25-test table, the smaller model wins. Chartography — reading values off charts — gives ZwZ2-9B a score of 13.00 and ZwZ2-27B a score of 12.00. Both are low. Neither is good. But the direction is the interesting part, because these two models are the same experiment at two sizes: one training recipe, one dataset, one paper-method, applied to a Qwen3.5-9B body and a Qwen3.6-27B body respectively. If the recipe is identical and the larger model is worth three times the parameters, the arithmetic of where that size actually shows up is more surprising than the headline overall scores suggest.

It is also worth saying at the top that both models arrived in the same quiet way. ZwZ2-9B's repository was created at 03:42 UTC on October 16, 2026; ZwZ2-27B's at 05:42 the same morning. Both carried weights and no documentation for ten days, and both received model cards with full benchmark tables on October 26. Neither declares a license. Neither has a published paper — the cards still contain the placeholder PAPER_URL.

One recipe, two bodies

The training method is not just similar between the two, it is the same method applied to two backbones, which makes them unusually clean to compare. EvidenceConnect groups three to seven related regions from SA-1B images, generates questions using privileged zoomed crops, filters those questions by cross-model agreement and a drop-one-region test, converts survivors into four-option tasks, and trains with GRPO — no supervised fine-tuning stage. Both models use approximately 20,000 training examples.

The bodies differ in the ordinary way:

• Scale — 9B parameters over 32 layers at 4,096 hidden size, versus 27B over 64 layers at 5,120 hidden size.

• Attention — 16 query heads and 4 KV heads at dimension 256 for the 9B, against 24 query heads and 4 KV heads at the same dimension for the 27B.

• Feed-forward — 12,288 intermediate size versus 17,408.

• Vision — the same tower in both: 27 layers at 1,152 hidden size, projecting to 4,096 and 5,120 respectively.

• Context and vocabulary — identical at 262,144 positions and 248,320 tokens.

• Multi-token prediction — removed from both, with mtp_num_hidden_layers at 0.

• Base models — Qwen3.5-9B for the small one, Qwen3.6-27B for the large one. That is the detail that complicates the comparison, and the next section is about why.

Where the 3.92-point overall gap actually comes from

Here is the finding I did not expect before doing the subtraction. The overall means are 60.51 for ZwZ2-27B and 56.59 for ZwZ2-9B — a gap of 3.92 points. Now look at what each recipe added to its own base, which is published in the same tables, by the same lab, on the same harness:

• ZwZ2-27B — base Qwen3.6-27B at 55.52, tuned at 60.51. Gain of 4.99.

• ZwZ2-9B — base Qwen3.5-9B at 51.63, tuned at 56.59. Gain of 4.96.

The recipe produces almost exactly the same improvement at both scales. Three times the parameters bought 0.03 points of additional training benefit. And the base models were already 3.89 points apart before either was touched.

So of the 3.92-point gap between the finished models, 3.89 points were inherited from the checkpoints they started from. The specialization did not create the gap. It preserved it almost perfectly.

This is a single-lab, single-harness comparison, so the subtraction is legitimate in a way that cross-vendor comparisons never are — every one of those four numbers comes from the same table, produced by the same evaluation code, on the same benchmark set. The rounding is coarse enough that 0.03 should not be read as precision; the honest reading is “indistinguishable.”

What it means practically is that choosing between these two models is very nearly choosing between Qwen3.5-9B and Qwen3.6-27B, plus a fixed perceptual adjustment. If you would not have picked the bigger base model for your workload, the ZwZ2 training is not going to change that decision.

Domain by domain: the size shows up in one place

Averages hide the structure, and here the structure is lopsided. Comparing the two tuned models across the six domains:

• Fine-grained perception — 67.57 for the 27B against 66.02 for the 9B. A gap of 1.55, the smallest of the six.

• Spatial perception — 55.98 against 52.33. A gap of 3.65.

• General visual reasoning — 53.71 against 49.83. A gap of 3.88.

• Embodied spatial reasoning — 55.71 against 51.11. A gap of 4.60.

• Long-context understanding — 68.77 against 64.08. A gap of 4.69.

• Multimodal agents — 51.81 against 42.43. A gap of 9.38.

The pattern is consistent: the bigger model's advantage roughly doubles when the task moves from looking at an image to acting on one. On the eight fine-grained perception benchmarks — the ones the method was explicitly designed for — the two models are within a point and a half of each other, with Chartography running backwards.

The agentic domain is where the 27B earns its memory footprint, and most of that gap comes from a single test. UI-Vision scores 66.66 for ZwZ2-27B and 45.48 for ZwZ2-9B — a 21.18-point separation, larger than any other in the table. ScreenSpot-Pro and AgentVista add smaller contributions, at 70.59 against 66.98 and 18.18 against 14.83.

UI-Vision is an interface-grounding benchmark, and the result says something specific rather than general: if your workload is reading and operating a software interface, the 27B is doing something the 9B mostly cannot. If your workload is comparing marked regions on a photograph or a document page, the 9B is doing nearly the same thing for a third of the weights.

The one test the small model wins, and how much it should count

Chartography is worth dwelling on precisely because it is the exception that proves how narrow the specialization is. Both scores are terrible — 13.00 and 12.00 out of 100, against a Qwen3.6-27B base of 8.00 and a Qwen3.5-9B base of 7.00. The training lifted both models by about five to six points on a task they still cannot perform reliably, and the 9B's lift happened to land one point higher.

The correct conclusion is not that the 9B is better at charts. It is that on a task this far outside the training distribution, the difference between the two models is noise, and the reasonable behaviour is to route chart-reading somewhere else entirely. A model scoring 12 or 13 on a chart benchmark does not have a chart capability to choose between.

The same caution applies to the other low absolute scores. ZwZ2-9B averages 42.43 in multimodal agents and 18.18 on AgentVista; ZwZ2-27B averages 51.81 and 18.18. Both are weak at agentic work in absolute terms, and the fact that the larger one is much less weak does not make it strong.

What the memory footprint really looks like

The parameter difference is the part you pay for every day. ZwZ2-9B is a single 9.41-billion-parameter safetensors file, roughly 19 GB in bf16. ZwZ2-27B is split across two shards at about 54 GB. Add vision-tower activations and a 262,144-position KV budget and the practical split is:

• ZwZ2-9B — a comfortable fit on a single 24 GB consumer accelerator at bf16, tight but workable; no quantized release exists, so you are carrying full precision unless you quantize yourself.

• ZwZ2-27B — an 80 GB card with headroom, or two 40 GB cards across a tight tensor-parallel split. Nothing below that runs it at bf16.

• Serving support — neither repository ships a vLLM or SGLang recipe. Both pin transformers==5.13.1 in the quickstart, and both are documented only in non-thinking mode with answers wrapped in <answer> tags. Neither declares a license, which for a production deployment is a bigger obstacle than either memory number.

That last point deserves its own sentence, because it is the same on both sides: there is no license to read. The cards defer to “the license accompanying the released weights,” and the weights carry none. Whichever size you pick, the commercial question is unanswered.

Using both, rather than choosing

The domain table suggests a two-tier pattern more strongly than it suggests a winner. Fine-grained perception barely moves between the sizes, so the 9B is a sensible first pass there; agentic interface work moves by more than nine points on the domain average and twenty-one on UI-Vision, so that traffic belongs on the 27B. A router that sends by task type would spend most of its tokens on the cheap model and reserve the expensive one for screens.

Neither ZwZ2 size is in our catalogue, so that pattern today means running two checkpoints on your own hardware and writing the dispatch yourself. Where OrcaRouter fits is the generalist side of the same stack: Qwen3.8-27B and Qwen3.5-27B are both routable here behind one key at provider list price with no markup, so the workload that does not need region comparison can go to a hosted model with automatic failover while the specialized traffic stays local. That is a smaller claim than “route ZwZ2 through us,” which we cannot offer — but it is the part of the architecture that is actually available now.

The honest summary is unglamorous. Two checkpoints, same recipe, same silence, same missing license, and a performance gap that is almost entirely the base models' pre-existing gap plus one real specialization at the top end for interface work. If you need the 27B's UI numbers, you need the 27B and the 80 GB card. If you do not, the 9B gets you within a point and a half on the task the whole project was built to improve, and it fits on hardware you probably already own.