Xing 4.0 29B-A4B vs Qwen 3.8 27B: Vision Versus Tool Calling

The two models are close enough in size that a procurement spreadsheet would put them on the same row: Qwen 3.8 27B is A​libaba's dense 27-billion-parameter multimodal model, 28B counting the vision tower, released under Apache 2.0 in August 2026 with 262,144 native positions and native image and video input; Xing 4.0 29B-A4B is China Telecom's sparse 29-billion-parameter mixture-of-experts model, 4B active per token, text in and text out, whose weights were published on 16 September 2026 and which only became callable in mid-October. Both are Apache 2.0, both are open-weight, and both report numbers that would have looked like frontier results a year ago. They are also not substitutes for each other in any workload that matters, and the reason is a single line in each specification: whether there is a vision tower on one side and a tool-calling loop on the other.

The size difference is smaller than the names suggest

• Naming — "27B" is a round label. The Q​wen repository reports 28B in BF16 tensors once the vision encoder is counted; the text backbone is 27B across 64 layers at a hidden size of 5,120 with a 248,320-entry vocabulary.

• Xing's labelling — "29B-A4B" is two figures, not one: 29B total and 4B active. The repository reports 31B in full-precision tensors.

• What you actually pay for — Qwen 3.8 27B is dense, so all 27B runs per token. Xing 4.0 reads 4B of 29B per token. At identical hardware and batch size, Xing is dramatically cheaper per generated token and Q​wen is dramatically more predictable.

• Context — Qwen 3.8 27B carries 262,144 native positions, extensible to 1,000,000, with a recommended ceiling of 262,144 reasoning tokens and 131,072 for the final response. Xing 4.0 carries 256K, claimed extensible to 512K. Comparable in the default configuration; Q​wen's published extension goes twice as far.

Architecture: two bets that do not resemble each other

Qwen 3.8 27B is built on a hybrid attention layout — sixteen blocks of three Gated DeltaNet layers followed by one full Gated Attention layer, a three-to-one split of linear attention against quadratic — and it is trained with multiple MTP steps. The stack is dense and linear-attention-heavy, which is what lets a 27B dense model carry a 262K window without the KV-cache bill a standard transformer would incur.

Xing 4.0 29B-A4B is a sparse mixture-of-experts model in the D​eepSeek lineage: MLA compressed-latent attention throughout, 64 routed experts with 4 active per token plus a shared expert, and manifold-constrained hyper-connection residual streams — a four-stream residual layout that serving frameworks need custom kernels for. It is also MTP-trained, so both models carry their own speculative-decoding head in the checkpoint rather than shipping a separate drafter file.

The consequence for a practitioner is unglamorous and important. Qwen 3.8 27B is a dense model with unusual attention, which most serving stacks handle with modest work. Xing 4.0's confirmed support is vLLM, SGLang and KTransformers, with SGLang carrying a kernel written specifically for the four-stream hyper-connection layout. A bespoke residual-stream design is a real integration cost the first time, and a recurring one every time you quantise, fine-tune, or change your batch shape.

What each model is actually for

Here the comparison stops being a comparison, because the two models were not built for the same job and their benchmark tables do not overlap enough to fake it.

• Software engineering — Xing 4.0 SWE-bench Verified 75.00 against Qwen 3.8 27B's SWE-bench Pro 61.7. These are different benchmarks, so the gap between them means nothing on its own. The comparable row is that Q​wen reports QwenSWEBench 79.0 and DeepSWE 1.1 42.2, which Xing's card never touches.

• Terminal and agentic coding — Xing 4.0 Terminal-Bench 2.1 57.50 against Qwen 3.8 27B Terminal Bench 2.1 (Terminus) 73.0. Same benchmark name, different harness detail, both vendor-reported. On the published figures Q​wen is ahead by more than 15 points.

• Computer use and vision-language agents — Qwen 3.8 27B OSWorld-Verified 84.3, WebArena-Verified 64.8, AndroidWorld 81.9, MathVision 90.0 without chain-of-thought and 94.6 with it, OmniDocBench 1.5 91.1, RealWorldQA 85.9. Xing 4.0 has no multimodal benchmark at all because it has no multimodal capability. This is not a gap in the table; it is a gap in the product.

• Tool calling and long-horizon agents — Xing 4.0 Claw-Eval 76.55, Tau3-Bench 64.63, DeepresearchBII 60.80. Qwen 3.8 27B CoWorkBench 70.7, the Agents' Last Exam at 20.4 pass@1 and 42.9 score, JobBench 33.4. Again: different suites, no overlap, no comparison available.

• General reasoning and knowledge — Qwen 3.8 27B GPQA Diamond 89.2, HLE 30.8, LiveCodeBench v6 90.3, IFBench 79.5. Xing 4.0's card reports only IFBench 69.67 on that list — nearly ten points behind Q​wen, and the one general-purpose row where both vendors published a figure.

Read the tables as two different documents about two different products. Q​wen's covers a multimodal model that does computer-use work through a screenshot interface. Xing's covers a text model built to hold a long tool-calling loop together. The single row where they genuinely collide — IFBench, instruction following — goes to Q​wen by 9.8 points.

Two sourcings, one shared caveat

Neither model has an independent evaluation. Qwen 3.8 27B's numbers are A​libaba's, published on the model card with harness settings, and Xing 4.0's are China Telecom's, published the same way. In both cases the comparison columns in the tables are the home lab's own re-runs of rival models rather than those rivals' published figures — a practice that is close to universal and worth remembering every time a vendor's table shows a rival losing badly.

Qwen 3.8 27B does have something Xing 4.0 lacks, and it is not a benchmark row: eight weeks in the open, including a community llama.cpp port reported running the weights on a jailbroken PlayStation 5's GPU at around 20 tokens per second. That kind of third-party work is the closest thing open weights have to an independent audit, and Xing 4.0 has had a month of it so far, most of which went into serving-framework pull requests rather than evaluation.

Cost, self-hosted and metered

Both are Apache 2.0, so the download price is zero and stays zero on both sides. Qwen 3.8 27B is dense and needs the memory of a 27B dense model plus a vision tower; the practical floor is a 24GB card at low precision, with a 55.6GB full-precision download. Xing 4.0's collection claims a 4-bit quantisation in roughly 15GB of VRAM — a claim worth checking against the actual GGUF file size, since the full-precision repository reports 31B in tensors.

On the metered side, Qwen 3.8 27B is the one that is actually callable here. OrcaRouter hosts qwen/qwen3.8-27b at $0.33 per million input tokens and $2.40 per million output, with the open weights served on our own infrastructure, which is why that price holds and why a vendor-side change to it would not. Qwen 3.8 27B is also available at $0.40 and $4.21 per million through a community FP8 deployment in our catalogue that keeps the vision tower at full precision — a useful reminder that two listings of the same open model are not the same product. Xing 4.0 is not in our catalogue, so the vendor's own API and third-party platforms are where you look, and we are not going to pretend otherwise.

Where a router is worth having in a comparison like this is on the decision itself. One API in front of 200-plus models at 0% markup means trying a challenger is a routing rule rather than a second integration, and automatic failover means the trial never puts the production path at risk: the primary answers, the fallback absorbs anything the primary drops, and the whole experiment is deleted from the config when it is over. Model fusion goes one step further and lets a panel of models answer together when a single model's confidence on a task is the thing you do not trust.

How to decide

Take Qwen 3.8 27B if anything in your pipeline involves an image, a screenshot, a scanned document, a video frame, or a computer-use loop. It is the only one of the two that can do that work at all, its computer-use benchmark numbers are the strongest published in this weight class, and it is callable on a metered basis today. It is also the safer default on instruction following, which is the failure mode that costs the most in a customer-facing pipeline.

Take Xing 4.0 29B-A4B if your workload is pure text, your workload is a long tool-calling loop, and you can afford to verify the vendor's claims yourself. The throughput-per-card argument is real on paper: a 29B model's memory footprint with a 4B model's arithmetic, at a licensing term your legal team already knows. The uncertainty is equally real — no independent evaluation, no general-reasoning rows published, and an instruction-following score almost ten points behind its competitor on the one benchmark both labs chose to report.

For most teams the answer is not a choice at all. Qwen 3.8 27B handles everything with a picture or a screen in it; Xing 4.0 handles the text-only agent loops where per-token arithmetic is the constraint. Running both through one endpoint is cheaper than deciding between them on evidence neither vendor has published, and the first independent evaluation of Xing 4.0 is the event that would turn this into a real comparison rather than an inventory of what each lab chose to measure.

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.