Xing 4.0 29B-A4B vs Nemotron 3.5 Lightning 30B-A3B: Same Shape, Different Licence

Put Xing 4.0 29B-A4B and N​emotron 3.5 Lightning 30B-A3B on the same line and the specifications almost converge: 29 billion parameters against 30 billion, four billion active per token against three, a sparse mixture-of-experts layout on both sides, and an explicit design target of agentic and tool-using work in both model cards. They even arrived five weeks apart in the same summer — N​VIDIA's general-availability release on 11 August 2026, China Telecom's weights on 16 September 2026, with a callable endpoint following in mid-October. The convergence is real and it makes this the most directly comparable pairing in this series. It is also where the resemblance ends, because the two releases make opposite bets on context length, on attention architecture, and above all on licensing — and one of those three is a decision your legal team makes, not your ML team.

Where the specs genuinely match

• Total parameters — N​emotron 3.5 Lightning 30B total vs Xing 4.0 29B. A one-billion-parameter difference on a thirty-billion model is nothing.

• Active parameters — N​emotron 3B per token vs Xing 4B per token. This is the number that decides your throughput, and the gap is far smaller than the naming difference ("A3B" against "A4B") makes it look: roughly a third more arithmetic per token for Xing.

• Form — both sparse MoE, both Apache-style fully open artefacts in the sense that the weights are downloadable and the serving recipes are published, both with quantised builds from the vendor and the community.

• Design brief — both cards are written for agents. Xing 4.0's says multi-step planning, tool calling and reasoning chains; N​emotron's ships alongside NeMo Switchyard, an open-source routing library for deciding which model handles the next step inside an agent loop. Two labs, the same conclusion about where small open models earn their keep.

• Publication — N​emotron 3.5 Lightning reached GA on 11 August 2026, with NVFP4 and base variants appearing across the following days. Xing 4.0's repository was created 16 September 2026, its quantised builds landed 13 October, and a metered endpoint appeared 15–16 October.

Where they stop matching: context, and how you get it

This is the largest practical gap between the two, and it is not close.

• Context window — N​emotron 3.5 Lightning: up to 1M tokens. Xing 4.0: 256K native, card-claimed extensible to 512K.

• What that means in practice — N​VIDIA's own card is candid that the million-token window is memory-bound and that 256K is the validated figure on a single 80GB H100. So the honest comparison at the deployment level is 256K against 256K, with N​emotron holding a documented, vendor-validated route to four times that and Xing holding a card claim to two times it.

• Attention design — N​emotron 3.5 Lightning is a hybrid: Mamba-2 state-space layers mixed with attention and MoE routing. Xing 4.0 is MLA throughout — compressed latent key-value state, the D​eepSeek-lineage approach — plus manifold-constrained hyper-connection residual streams. These are two different strategies for the same KV-cache problem, and neither is a mistake.

• Speculative decoding — Xing 4.0 has multi-token prediction trained into the checkpoint. N​emotron 3.5 Lightning's published artefacts include draft-model variants in its quantised family, which serve a similar purpose, but MTP is a property of Xing's weights rather than a companion file.

• Precision — N​VIDIA publishes the BF16 checkpoint as a full-precision reference explicitly intended for customization and post-training, with NVFP4 builds for production inference. Xing 4.0 publishes BF16, FP8, w8a8 and GGUF. If you plan to fine-tune, the two are equally well served; if you plan to serve, both want you on the quantised artefact.

The licensing difference is the real decision

N​emotron 3.5 Lightning ships under the OpenMDW License Agreement, version 1.1 — N​VIDIA's own open-model licence, stewarded outside the company. Xing 4.0 ships under Apache 2.0.

Both permit commercial use. They are not equivalent documents, and the difference shows up in the parts of a deployment that engineers rarely read. Apache 2.0 is a mature, widely litigated, OSI-approved licence with a familiar patent grant and a decade of corporate legal precedent behind it; it is pre-cleared at most large organisations. OpenMDW is newer and less familiar, which means the first commercial deployment of a N​emotron model inside a regulated or enterprise environment is more likely to route through legal review than the equivalent Xing 4.0 deployment would. That is not an argument that OpenMDW is restrictive — it is an argument that unfamiliarity has a cost, and it is a cost paid in weeks rather than dollars. If you are choosing between these two on technical merit alone, you are ignoring the half of the decision your legal team owns.

Both benchmark tables are vendor tables

This is the part of the comparison where a reader should slow down. Neither model has an independent evaluation on a harness chosen by someone other than the vendor.

• SWE-bench Verified — Xing 4.0 75.00 vs N​emotron 3.5 Lightning 51.56.

• SWE-bench Multilingual — Xing 4.0 66.00 vs N​emotron 3.5 Lightning 39.33.

• Terminal-Bench 2.1 — Xing 4.0 57.50 vs N​emotron 3.5 Lightning 24.58.

• Instruction following — Xing 4.0 IFBench 69.67 vs N​emotron 3.5 Lightning IFBench 71.88, reported on the loose setting.

• General reasoning — N​emotron 3.5 Lightning MMLU Pro 81.94, GPQA Diamond 75.44, HLE 11.72, SciCode 32.60, AA-Omniscience 17.50. Xing 4.0's card publishes no comparable row.

• Long context — Xing 4.0 AA.LCR 61.00 vs N​emotron 3.5 Lightning AA-LCR 52.00. Xing wins the one long-context row where both numbers exist, which is a mildly surprising result given N​emotron's much larger stated window.

• Agentic browsing and retrieval — N​emotron 3.5 Lightning BrowseComp 36.97, PinchBench 85.37, τ³-bench (Banking) 9.28, GDPval-AA-V2 832. Xing 4.0 Claw-Eval 76.55, Tau3-Bench 64.63, DeepresearchBII 60.80. These suites do not overlap at all, so no comparison is possible and any table that lines them up is inventing one.

The provenance behind them differs in one interesting way. N​VIDIA states its results were measured under a consistent internal harness and warns that they "may differ from vendor-reported figures" — and its published table includes competitor columns for Qwen 3.6 35B A3B, Gemma 4 26B A4B and G​PT-OSS 20B, which are N​VIDIA's own re-runs of those models. China Telecom's table does the same thing in the other direction, benchmarking itself against its own re-runs of G​emma 4 and Q​wen. Each lab is measuring its rivals in its own harness and publishing the result as a comparison column. Neither is doing anything dishonest, and neither column should be read as the rival's own figure.

Cost and serving

N​emotron 3.5 Lightning's 1M-token window has a real operational consequence that engineers discover the hard way: it is memory-bound, and a request that exceeds available KV-cache capacity fails at the provider rather than degrading gracefully. That is the shape of failure a router is built to absorb — OrcaRouter retries against healthy fallback capacity before the response begins, so an upstream out-of-memory or a 429 on a long-context call does not surface to the user as an error. We route 200-plus hosted models behind one API at 0% markup, provider list price passed through with nothing added, and this is the failure mode where the failover chain stops being a nice diagram and starts being the reason your agent finishes its task.

Neither N​emotron 3.5 Lightning nor Xing 4.0 is in our catalogue today, so we are not a route to either model and this piece will not imply otherwise. What the numbers look like if you run them yourself: Xing 4.0's 4-bit quantisation is claimed at roughly 15GB of VRAM, which would put it on a single RTX 3090 or 4090; N​emotron 3.5 Lightning's memory-bound behaviour on a single 80GB H100 is documented for its validated 256K window, which is the more demanding target of the two. Xing 4.0 is the cheaper model to host; N​emotron 3.5 Lightning is the one with a documented path to a much longer context if you have the hardware for it.

How to choose

Take N​emotron 3.5 Lightning 30B-A3B if context length is the binding constraint, if you are already inside N​VIDIA's stack and NeMo Switchyard is a natural fit for your agent loop, or if your organisation has already cleared OpenMDW for commercial use. The wider window, the published base-model route for customization, and the fact that its evaluation table covers fourteen benchmarks rather than nine all point the same direction.

Take Xing 4.0 29B-A4B if licence clearance speed matters, if your work is software engineering and terminal-shaped agent tasks rather than long-document retrieval, or if the throughput-per-card argument is what you are actually buying. Its agentic margins on its own table are the largest in this weight class, its licence is the one every lawyer already knows, and its long-context row — the single place both models published a comparable figure — went to Xing.

What would change the answer is the same thing that would change it in every comparison in this series: one independent evaluation of both models, run by a third party, on one harness. N​VIDIA and China Telecom have each published a table that flatters their own model against re-runs of the other's generation. When someone publishes the first table neither of them wrote, the 23-point SWE-bench gap above becomes either the most important number in this weight class or an artefact of two different agent harnesses. There is no way to tell which from the outside.

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.