ZwZ2-27B: The Weights Landed on October 16, the Explanation Arrived Ten Days Later

ZwZ2-27B has been downloadable from Hugging Face since October 16, 2026. For the first ten days of its public life there was no announcement, no paper, no README and no stated license — just two bf16 safetensors shards under inclusionAI, and a repository that gave no reason for existing. On October 26 a model card appeared. It says the model is a reinforcement-learning-trained vision-language system built on Qwen3.6-27B, and it publishes a 25-benchmark evaluation in which the model beats that base by 4.99 points overall.

That is the whole of what is new, and it is worth separating from what people will assume happened. This is not a launch. The weights are eleven days old and the vendor still has not announced them. What changed this week is that we can finally read what inclusionAI thinks the model is for.

What actually changed, by date

The repository's own commit history is the cleanest evidence available, and it does not flatter the story. Three commits carry everything:

• October 16, 05:42 UTC — initial commit, then “Add ZwZ2-27B model weights” ten minutes later. Config, tokenizer, processor and the two weight shards. No README.

• October 26, 02:58 UTC — “Create README.md”. The first written description of the model, ten days after the weights.

• October 26–27 — repeated README revisions and six-file uploads as the benchmark figures and image assets were assembled.

At the time of writing the repository shows seven downloads and zero likes. The companion GitHub project, inclusionAI/Zooming-without-Zooming-v2, was created on October 14 and has one star. Searches for the model name in English return coverage of the older ZwZ-2B/4B/8B family and of the October 16 ZwZ2-9B, and nothing at all about the 27B. If a launch happened here, it happened without an audience.

The model in the config file

Config files are the least glamorous and most reliable part of any release. This one describes a dense vision-language model whose architecture declaration is Qwen3_5ForConditionalGeneration — the same class Qwen's own models use — with the following numbers:

• Text tower — 5,120 hidden size, 64 layers, 24 query heads and 4 KV heads at head dimension 256, feed-forward width 17,408.

• Attention layout — a 3:1 ratio of linear-attention layers to full-attention layers, repeating 16 times, with interleaved M-RoPE and a rotary base of 10,000,000.

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

• Vision tower — 27 layers at 1,152 hidden size, projecting to 5,120, patch size 16 with 2×2 spatial merge.

• Multi-token prediction — mtp_num_hidden_layers set to 0, i.e. the MTP head has been removed relative to the Qwen backbone.

That last line is the only structural difference from the base. Everything else — layer count, head geometry, vocabulary, context ceiling, vision depth — matches Qwen3.6-27B exactly. ZwZ2-27B is the same body with a different post-training regime and one speculative-decoding head taken off.

It also means the model is not small. At 27B parameters in bf16 the weights are roughly 54 GB across two shards, which puts a comfortable load at a single 80 GB accelerator or a tight one across two 40 GB cards. There is no quantized variant in the repository, no vLLM or SGLang recipe, and no serving notes beyond a Hugging Face transformers quickstart pinned to version 5.13.1.

The +4.99 claim, and exactly what it covers

The card's headline is that across 25 benchmarks in six domains the model “improves over its base model on every benchmark,” moving the overall mean from 55.52 to 60.51. Both halves of that sentence check out against the published results figure, and both need a qualifier.

The qualifier is that this is inclusionAI's own harness, its own benchmark selection, and its own scoring, published on its own repository. Nothing here has been reproduced outside the organisation. It is a vendor evaluation in the ordinary sense: useful for seeing what the team optimised for, not usable as an independent grade.

The other qualifier is scale. A 4.99-point gain on a 25-benchmark mean is real movement, but the mean is an average of accuracies across very different tasks, weighted equally by domain. It is not a capability index and should never be compared with a score from a different suite.

Here is the domain-level breakdown, base versus tuned:

• Fine-grained perception — 62.09 to 67.57, a gain of 5.48 across eight benchmarks including VStar, HR-8K and MME-RealWorld.

• Spatial perception — 49.72 to 55.98, a gain of 6.26 across six benchmarks.

• Long-context understanding — 65.87 to 68.77, a gain of 2.90 across three benchmarks.

• Embodied spatial reasoning — 50.17 to 55.71, a gain of 5.54 across three benchmarks.

• Multimodal agents — 49.73 to 51.81, a gain of 2.08 across three benchmarks.

• General visual reasoning — 47.89 to 53.71, a gain of 5.82 across MMStar and BabyVision.

Where the gains are concentrated — and where the model is still weak

The shape of that table matters more than its average. The largest relative gains land in spatial perception and embodied reasoning, which is consistent with what the method claims to teach: comparing, matching and aggregating evidence across several marked regions of one image, using full images at inference time and no external zoom calls.

The two smallest gains are long-context and multimodal agents. Those are also the two domains where the base model was already strongest relative to the task, which is the ordinary pattern for RL fine-tuning on a narrow data recipe: it moves what the reward signal can see.

The absolute numbers are the more sobering read. Chartography sits at 12.00. AgentVista sits at 18.18. Perception-Bench at 44.93. A model can gain five points overall and still be close to unusable on several individual benchmarks in the set, and this one is. Anyone quoting “60.51 overall” as a general competence figure is quoting an average that hides a 12.

The training recipe explains the ceiling. EvidenceConnect builds its data from SA-1B images by grouping three to seven related regions, generating questions with privileged zoomed crops, filtering through cross-model agreement and drop-one-region tests, converting to four-option tasks, and training with GRPO — all three model sizes on roughly 20,000 examples, with no supervised fine-tuning stage. That is a focused diet. It reliably buys region-level perception and does not pretend to buy open-ended reasoning.

The reference-model columns deserve a warning

The results figure also carries three API models as reference points: GPT-5.6-Luna at 54.02, GLM-5.3-Flash at 57.24 and DeepSeek-V4.1-Flash at 55.47, against ZwZ2-27B's 60.51. On its face that puts a quietly uploaded open-weights checkpoint above three hosted production models.

Three things should temper that reading. First, the serving constraint is noted in the card itself: advanced models' image resolution was constrained by their serving platforms, which is a real handicap on a benchmark set that is heavy on fine detail. Second, these are not those vendors' own published evaluations — they are inclusionAI running other companies' models through inclusionAI's harness, with no indication that the vendors were involved or have confirmed the figures. Third, the GPT column is explicitly marked as not reported for some rows, so the reference set is unevenly populated.

None of that makes the numbers wrong. It does mean they are one lab's measurements of its competitors on a benchmark suite where its own model had ten days of extra tuning. Treat the base-model comparison as the meaningful one and the API columns as context.

What is still missing

The honest inventory of unknowns is longer than the list of knowns, and it starts with the license. The model card's front matter tags the base model, the dataset and the pipeline task; it does not declare a license. The body says to “refer to the license accompanying the released weights and the applicable backbone license,” which is circular when the weights themselves do not carry one. Qwen3.6-27B is Apache-2.0, so the backbone side is clear. The derivative's terms are not, and for a company that wants to ship this in a product, that single omission is more consequential than any benchmark row.

The paper is also not out. The card's own header contains the literal placeholder text <!-- Replace PAPER_URL with the final v2 paper URL. --> and links to PAPER_URL. The project repository exists; the write-up it should point to does not.

Beyond that:

• No independent evaluation of ZwZ2-27B exists, from anyone.

• No vendor announcement, blog post or social post has been located for this model.

• The published evaluation is low/thinking mode — the card's quickstart runs with enable_thinking=False and the chat template asks for answers inside <answer> tags. Thinking-mode numbers are not part of the claim.

• No framework support beyond the pinned Transformers version. No GGUF, no AWQ, no published vLLM config.

• The card lists known failure modes — hallucinated detail, misread text, counting and spatial errors — which is more candour than most vendor cards offer and a fair summary of what a 12.00 on Chartography looks like in practice.

How to evaluate it without betting anything on it

ZwZ2-27B is not in our catalogue, and this piece should not be read as implying it is. To run it today you download 54 GB of weights, install a specific Transformers build, and serve it yourself — which is a reasonable afternoon if you already have the hardware and a poor use of one if you are trying to answer a product question this week.

A cheaper sequence is to test the claim rather than the model. The claim is narrow: that RL on a 20K-example region-comparison dataset makes a Qwen3.6-27B checkpoint better at reading fine detail without zoom tools. If that is your actual bottleneck — dense documents, charts, UI screenshots, multi-region questions — the base model is the control condition, and its successor is already routable here: Qwen3.8-27B and Qwen3.5-27B are both on OrcaRouter behind one key, alongside 200-plus other models at provider list price with no markup. If a self-hosted ZwZ2-27B turns out to be worth adopting, you can put it on the same key and let automatic failover decide between it and a hosted generalist rather than rewriting the integration.

What would move this from “interesting upload” to “adoptable model” is a short list. A named license. A paper that lets someone check the benchmark construction. One independent run — even a partial one — on benchmarks the vendor did not choose. And an announcement, which after eleven days of silence is starting to look like a deliberate absence rather than an oversight.

Until then the defensible statement is narrow and worth stating plainly: an undocumented 27B checkpoint from inclusionAI has been sitting on Hugging Face since October 16, and as of October 26 its authors are willing to claim a five-point improvement over Qwen3.6-27B on a benchmark suite they built themselves. That is a hypothesis with weights attached. It is not yet a result.