Xing 4.0 29B-A4B vs Gemma 4 26B-A4B: Two Sparse Models, One Has an API

Both of these models are sparse mixture-of-experts checkpoints that activate roughly four billion parameters per token, and both are Apache 2.0. That is where the resemblance stops being useful. Xing 4.0 29B-A4B is China Telecom's text-only, agent-oriented MoE — 29B total, 4B active, weights published on 16 September 2026 and given a callable commercial endpoint only in mid-October. Gemma 4 26B A4B is G​oogle DeepMind's multimodal sparse model — 25.2B total, 3.8B active on the instruction-tuned checkpoint, weights public since 11 March 2026 — and it has been available through hosted inference at six cents per million input tokens for months. If you are choosing between them, the honest framing is not "which is smarter" but "which one can you actually put in a production path this week, and what do you give up for that."

The one-line answer

Gemma 4 26B A4B, unless your workload is a coding or terminal agent running on your own hardware. Xing 4.0's entire pitch is agentic strength per active parameter, and on China Telecom's own table it wins the agentic rows by enormous margins. G​emma 4 wins on availability, modality, price transparency, and the simple fact that its numbers were published by the company that built it while its rival's comparison column was not.

Specs, dimension by dimension

• Size — Xing 4.0 29B total / 4B active vs G​emma 4 25.2B total / 3.8B active. The name on the G​emma card is a rounded label; the actual figures are the ones G​oogle states.

• Expert layout — Xing 4.0: 64 routed experts, 4 active, plus 1 shared. G​emma 4: 128 total experts, 8 active, plus 1 shared. G​emma has the finer-grained routing; Xing has the sparser read.

• Attention — Xing 4.0 uses MLA, multi-head latent attention, with compressed key-value state. G​emma 4 uses hybrid attention: local sliding-window layers with a 1,024-token window interleaved with full global attention, and proportional RoPE. Different answers to the same KV-cache problem.

• Depth and width — Xing 4.0: 40 layers, hidden size 3,584, vocabulary 131,072. G​emma 4: 30 layers, vocabulary 262K.

• Context — Xing 4.0: 256K native, card-claimed extensible to 512K. G​emma 4: 256K. Effectively level, though G​emma's long-context behaviour has had eight months of third-party testing and Xing's has had none.

• Modality — Xing 4.0 is text in, text out. Gemma 4 26B A4B takes text, image and video in, and returns text; its vision encoder is roughly 550M parameters. This is not a small gap if your workload touches documents or screenshots.

• Licence — Apache 2.0 on both, including the patent grant. Genuinely equivalent.

• Publication date — Gemma 4 26B A4B: instruction-tuned weights on Hugging Face since 11 March 2026, with quantisation-aware builds through April and May. Xing 4.0: initial repository commit 16 September 2026, quantised builds 13 October, hosted listing 15–16 October.

• Hosted price — Gemma 4 26B A4B is listed on OrcaRouter at $0.06 per million input tokens and $0.33 per million output, with cached input reads at $0.0075 per million. Xing 4.0 has no comparable published rate on a platform we can point at; the mid-October listing carried a free promotional tier, and the vendor's own route is the other option.

The benchmark question, which is messier than it looks

China Telecom's model card benchmarks Xing 4.0 against a column labelled Gemma4-26B-A4B. That column reports SWE-bench Verified 53.00. G​oogle's model card for the instruction-tuned checkpoint does not report SWE-bench Verified at all — the published suite is MMLU Pro, AIME 2026, LiveCodeBench v6, Codeforces ELO, GPQA Diamond, Tau2, HLE, MMMU Pro, MATH-Vision and a long-context retrieval test. So the 53.00 in that table is not G​oogle's number. It is China Telecom's own run of G​emma 4 through an SWE-agent harness, and there is nothing wrong with that as long as you read it as what it is: one lab's re-run, with no stated attempt to tune G​emma's scaffolding.

With that caveat stated once and carried through everything below, here is the contrast on the rows where both sides have a published figure, plus the agentic rows where only one does.

• SWE-bench Verified — Xing 4.0 75.00 vendor-run vs G​emma 4 53.00 as re-run by China Telecom; G​oogle publishes no figure for this benchmark.

• Terminal-Bench 2.1 — Xing 4.0 57.50 vs G​emma 4 30.00, again the vendor's own re-run of G​emma.

• SWE-bench Multilingual — Xing 4.0 66.00 vs G​emma 4 51.00, same provenance.

• Claw-Eval — Xing 4.0 76.55 vs G​emma 4 71.49, official harness on both sides per the card.

• Tau3-Bench — Xing 4.0 64.63 vs G​emma 4 58.90. G​oogle reports Tau2 at 68.2 on its own card, which is a different version of the family and not directly comparable.

• AIME 2026 — Xing 4.0 90.00 vs G​emma 4 88.30.

• IFBench — Xing 4.0 69.67 vs G​emma 4 72.67. G​emma wins this one.

• AA.LCR — Xing 4.0 61.00 vs G​emma 4 66.00. G​emma wins this one too.

• Where only G​oogle has numbers — MMLU Pro 82.6, GPQA Diamond 82.3, LiveCodeBench v6 77.1, MMMU Pro 73.8, Codeforces ELO 1,718. Xing 4.0's card reports nothing on any of these, so there is no basis for a comparison and the absence should not be read as a loss.

The shape of the difference is consistent and it is not random: Xing 4.0 is strong on software-engineering and terminal-shaped tasks and weaker on instruction following and long-context reasoning. That is a deliberate design target — the card's whole architecture section is oriented around agentic work: multi-step planning, tool calling, reasoning chains — and the numbers agree with the design. What they do not tell you is whether that strength survives a harness that was not chosen by the vendor.

What the G​emma column cannot tell you, and Xing's cannot either

Note the asymmetry in scrutiny. Gemma 4 26B A4B has been downloaded more than ten million times from its repository and has an independent Artificial Analysis placement and eight months of practitioner reports about where the sliding-window attention degrades. Xing 4.0 has 53,000 downloads in a month, no independent evaluation, and a single lab's benchmark table — and its most impressive rows are the ones most sensitive to harness choice. When one side of a comparison has been stress-tested by strangers for eight months and the other has not, the newer model's headline numbers deserve exactly the discount that fact implies. The 75 and 57.5 might hold up perfectly. Nobody outside China Telecom knows yet.

Cost, both ways

Gemma 4 26B A4B's case is arithmetic. At $0.06 and $0.33 per million tokens on OrcaRouter, with a cached read at $0.0075 per million, an agent burning two million output tokens a day pays roughly sixty-six cents for generation and pennies for the input side, before caching. There is no self-hosting requirement, no hardware decision and no ops burden. Because OrcaRouter passes provider list price through at zero markup, a vendor price cut on this model lands on our endpoint the same day it is announced — the number in your bill is the provider's number.

Xing 4.0's case is capital rather than marginal. The weights are free and Apache-2.0, and the collection claims a 4-bit quantisation running in roughly 15GB of VRAM, which would put it on an RTX 3090 or 4090. That claim sits awkwardly beside the 31B tensor size the repository reports for the full checkpoint, so check the actual quantised file's size before you plan around it. Serving it means SGLang, vLLM or KTransformers on hardware you own, plus the integration work that comes with a bespoke attention and residual-stream design. The crossover only arrives at sustained high volume, or when the data cannot leave your network at all.

Where a router helps with a model we do not host

We will say the obvious thing plainly: Xing 4.0 29B-A4B is not in OrcaRouter's catalogue, so we are not a way to call it, and nothing here should be read as a pitch to use us for it. What we are is a way to keep the comparison cheap. Gemma 4 26B A4B is hosted here; a challenger evaluation is a routing change rather than a second contract, and automatic failover keeps the production path alive while you point a slice of traffic at something experimental. The routing DSL exists for exactly this shape of problem — you write the rule once, and the model behind it is a variable.

Which one to pick

Take Gemma 4 26B A4B if you need multimodal input, a published per-token price, a model with eight months of independent testing behind it, or simply something you can call this afternoon. It is the safer of the two by a wide margin, and at six cents per million input tokens the cost argument barely needs making.

Take Xing 4.0 29B-A4B if your workload is a coding or terminal agent, you have the GPU or the tolerance for a brand-new hosted endpoint, and you are willing to run your own evaluation before trusting the table. The potential payoff is real — a model with the memory footprint of a 29B and the compute profile of a 4B, posting the strongest agentic rows of anything in its weight class on paper. The risk is equally real: every one of those rows is a vendor's claim attached to a checkpoint four weeks old, and the comparison column next to it was produced by the same vendor.

If you want the middle path, it is a two-model stack rather than a choice: Gemma 4 26B A4B handling the multimodal, high-volume and long-context traffic it has already proven itself on, and Xing 4.0 restricted to the coding-agent path where its claimed advantage is largest and where a wrong answer is cheap to catch.

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.