ZoomGain-9B vs Qwen 3.8 8B: There Is No Qwen3.8-8B, and That Is the Article

Search for "Q​wen 3.8 8B" and you will find comparison pages, download links and half a dozen articles treating it as a shipping model. It is not one. A​libaba's Q​wen3.8 generation contains four open checkpoints — Qwen3.8-27B, its FP8 build, Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next — and none of them is an 8B. The 8B slot in the Q​wen line-up is occupied by Qwen3-8B, a model from the 2025 Q​wen3 generation, which is a different architecture, a different tokenizer and a different era. So the honest version of this matchup is between ZoomGain-9B, the 9B vision-language checkpoint inclusionAI published on 10 November 2026, and the two Q​wen models that actually exist at a comparable size: Qwen3.8-27B, the smallest dense member of the 3.8 generation at 27 billion parameters, and Qwen3.5-9B, the trunk ZoomGain-9B was post-trained from.

That correction is not a technicality to clear out of the way before the interesting part. It changes the answer. Qwen3.5-9B is the base model ZoomGain was trained on, so it is the closest thing to a controlled experiment in this entire release: same weights, same architecture, same tokenizer, 9,000 RL samples of difference. Qwen3.8-27B is three times the size and two generations newer, and it is the model you would actually be choosing instead of ZoomGain if you wanted a Q​wen endpoint today. Comparing against both tells you something the phantom 8B never could.

Why the phantom entry exists at all

The naming is not random. Q​wen shipped 27B and 2.4T-A95B on 5 and 8 August 2026, then Flash-Next on 24 August. The previous generation, Q​wen3.5, had a 9B and a 4B; Q​wen3.6 had 27B and 35B-A3B. Every generation since Q​wen3 has included a small dense model, and the 3.8 generation is the first to skip one. Anyone pattern-matching on the previous three releases would reasonably predict a Qwen3.8-9B, and a Qwen3.8-8B is one digit away from that prediction.

What actually fills the gap is community distillation. Qwen3.8-9B-Distill checkpoints appeared on the Hub from independent labs within a week of the 27B release, and derivatives of those derivatives followed. They are small models trained on Q​wen3.8 outputs, not Q​wen3.8 models. If a page you found is describing "Q​wen 3.8 8B" with benchmark numbers, the most likely explanation is that it is describing one of those, or that it is describing nothing.

The practical consequence: if you need a small Qwen-family model, the current real options are Qwen3.5-9B at 9B and Qwen3.5-4B at 4B, both from February 2026, both Apache 2.0, both still among the most downloaded models on the Hub. ZoomGain-9B is built on the first of them.

The controlled experiment that the release accidentally provides

Qwen3.5-9B's configuration and ZoomGain-9B's are the same shape: 32 layers, hidden size 4,096, intermediate size 12,288, 16 query heads against 4 KV heads at head dimension 256, 262,144 maximum positions, and the same 3:1 schedule of linear DeltaNet to full attention. The vision tower is the base model's. Nothing about the architecture changed.

What changed is the post-training. ZoomGain applied GRPO to 9,000 questions generated by mining cropped detail, for one epoch, with eight rollouts per question. That is the whole intervention, and its measured effect on ZoomGainBench is 32.13 to 55.48 — a 23.35-point jump on a benchmark built from the same pipeline as the training data, which the paper honestly labels "Held-in".

The number that survives that caveat is the held-out average across thirteen benchmarks the training never saw: 59.75 for Qwen3.5-9B, 66.10 for ZoomGain-9B. A 6.35-point gain from 9,000 samples, on the same weights. In the specific family the release targets, the deltas are larger: HR-Bench 4K 82.13 to 90.25, V*Bench 86.91 to 93.19, VisualProbe 44.66 to 56.31.

• The interesting negative result — MMStar moves from 71.73 to 73.87, and BabyVision from 15.98 to 17.53. General multimodal competence barely shifts. The training bought fine detail and not much else, which is what a 9,000-sample targeted RL run should do.

• The honest caveat — all of the above is author-reported on the authors' harness. Nobody outside Ant Group has reproduced it. The evaluator and the benchmark are public, so it is checkable, but it has not been checked.

Qwen3.8-27B, and the row that puts it in the same table

Here the comparison gets real, because the ZoomGain paper did evaluate Qwen3.8-27B on its own benchmark. The 27B scores 44.81 overall on ZoomGainBench — behind Kimi-K3's 48.65, behind Qwen3.8-Flash-Next's 55.05, and behind every closed-source frontier model in the table. It is, by that measure, a competent but unremarkable fine-detail perceiver.

Set that against ZoomGain-9B's 55.48 on the same test and the arithmetic looks striking: a 9B model scoring 10.7 points above a 27B model from the same family, on a benchmark neither vendor's marketing would choose. But this is exactly the comparison that needs the held-in label attached, and the paper attaches it. ZoomGainBench was built by the pipeline that built ZoomGain's training data. A model trained on questions from that pipeline has seen the distribution; a model evaluated on it has not. The 10.7-point gap is a statement about distribution fit, not about capability.

Where the two can be compared without that distortion, the picture inverts. Qwen3.8-27B's own card reports Terminal Bench 2.1 at 73.0, SWE-bench Pro at 61.7, LiveCodeBench v6 at 90.3, GPQA Diamond 89.2 and HLE 30.8. ZoomGain-9B was evaluated on none of these and would not be expected to do well: its training never touched code, terminal tasks or graduate-level science. Qwen3.8-27B also takes images and video natively, serves a 262,144-token context extensible to one million, and supports thinking control with `reasoning_effort` — a capability ZoomGain-9B does not have at all, since its evaluation runs with thinking disabled.

• Size and shape — Qwen3.8-27B is 27B dense with 64 layers, hidden 5,120, intermediate 17,408 and a 1-layer multi-token-prediction head. ZoomGain-9B is roughly 9.4B with 32 layers at hidden 4,096.

• Availability — Qwen3.8-27B has 7.4 million downloads, an Apache 2.0 licence, FP8 and GGUF builds, and a hosted endpoint. ZoomGain-9B has no hosted endpoint, an empty provider mapping on the Hub, a licence deferred to a card with no licence tag, and 11 November GGUF builds as its only distribution beyond BF16.

• Release dates — Qwen3.8-27B: 5 August 2026, a full launch with a blog post and a technical report. ZoomGain-9B: 10 November 2026, a paper and a repository and nothing else. The three-month gap matters less than the fact that one of them has been in production hands since August.

The price of each route, and the one we can actually sell you

Qwen3.8-27B is in the OrcaRouter catalogue at $0.33 per million input tokens and $2.40 per million output, with a 262K context, text, image and video input, and 0% markup on the provider's list price — the number on our model page is the number Q​wen charges, passed through, so a vendor price change is live here the same day rather than after a repricing cycle. The same key also reaches Qwen3.8-Flash and Qwen3.8-Max if the 27B turns out to be the wrong size, and failover means a stalled endpoint retries to a provider before the response starts rather than after your request times out.

ZoomGain-9B is not in the catalogue and cannot be called through us. It is 18.82 GB of BF16, or 5.63 GB at Q4_K_M plus a 0.52 GB mmproj projector, served under vLLM 0.19.1 with transformers 5.7.0 or under a llama.cpp build with Q​wen3.5 multimodal support. Qwen3.5-9B itself is also not in the catalogue, which is worth saying plainly: the cheap hosted version of this comparison does not exist on our side either.

If your actual question is "should I run ZoomGain-9B instead of paying for a Q​wen endpoint", the cost comparison is not close in either direction and does not need to be. A 27B dense model at $0.33 per million input tokens is cheaper than owning a GPU for almost any volume you would describe as occasional. Self-hosting a 9B only wins when the volume is high, the data cannot leave your network, or the task is the one narrow thing the 9B was trained for.

When the phantom model would have mattered

There is a real gap in the Q​wen line-up, and it is worth naming because it explains why so many people searched for this matchup. A 9B dense model with a 262K window and a Qwen3.8-generation vocabulary, sold at something near Qwen3.5-9B's price, would be the obvious budget choice for image-heavy pipelines — cheaper than the 27B, more capable than the 4B, and small enough to self-host on a single card. Q​wen did not ship one, so the slot is currently served either by the older Qwen3.5-9B or by a community distillation with no vendor behind it.

ZoomGain-9B is not that model either, and the distinction matters. It is a perception specialist with a research licence question and no hosted path, not a general-purpose small Q​wen. If what you wanted from "Q​wen 3.8 8B" was a cheap generalist, ZoomGain-9B will disappoint you on everything except fine detail. If what you wanted was a model that can find the serial number in a 4K photograph, it is one of the few small checkpoints built specifically for that.

For the generalist case, the realistic substitutes are the hosted Q​wen3.5 line — 35B-A3B at $0.057 per million input tokens, 27B at $0.086 — which occupy the budget slot the missing 8B would have filled, with a larger footprint and a real endpoint. For the perception case, there is no hosted substitute we can offer: Qwen3-VL-8B-Instruct at $0.18 input and $0.70 output per million tokens takes images and works as a stand-in to test whether your errors are detail errors, but it has not been trained the way ZoomGain was and will not match it on crops.

What to do with this, concretely

If you came here for a ZoomGain-9B vs Qwen3.8-8B verdict, the verdict is that the opponent does not exist and any page that gave you one invented half of the comparison. The two real questions are narrower and both answerable today.

Question one: is a 9B fine-tune of Qwen3.5-9B better than Qwen3.5-9B at fine-grained perception? On the authors' numbers, yes by 6.35 points held-out and by much more on the benchmarks the training targeted. That is a strong result for 9,000 samples and the most interesting thing in the release.

Question two: is ZoomGain-9B better than Qwen3.8-27B at anything that matters to you? On general capability, no — the 27B is newer, larger, better documented, licensed, hosted, and evaluated on far more axes. On a benchmark built from ZoomGain's own training distribution, it scores 10.7 points higher, which tells you the training worked and tells you nothing about which model to buy.

The decision that follows is unglamorous. Use a hosted Q​wen model for general work, because it costs $0.33 per million input tokens and arrives in one API call. Download ZoomGain-9B only if you have a specific fine-detail job, a GPU, and a lawyer who has read a licence file that the vendor has not published.