Nanbeige4.3-33B vs Qwen 3.6 35B-A3B: Nanbeige Benchmarked a Different Qwen, and That Is the Story

Open Nanbeige's model card for Nanbeige4.3-33B and count the Qwen columns. There is exactly one, and it belongs to Qwen3.8-27B — a dense multimodal model released in August 2026. Qwen 3.6 35B-A3B, the mixture-of-experts model that Liquid AI used as its own reference point for LFM2.5 and that most of the open-weight agent ecosystem benchmarks against, does not appear in Nanbeige's twenty-row table at all.

That absence is a choice, and it makes this matchup more interesting than a normal head-to-head. Nanbeige4.3-33B was compared to the Qwen model it could beat on long-horizon agentic rows. Qwen 3.6 35B-A3B is the Qwen model that costs a quarter as much to run, accepts images and video, and has been in production since April. Whether that is a fairer comparison or an easier one depends on what you are buying, and the two models are not optimised for the same buyer.

One of these is dense and the other is not, and that decides the bill

The single most consequential difference is not on any benchmark.

• Parameters — 33.1B total, all active, for Nanbeige4.3-33B. 35B total with roughly 3B active per token for Qwen 3.6 35B-A3B, a mixture of experts.

• Compute per token — a full dense forward pass for Nanbeige4.3-33B versus a sparse one over a small expert subset for Qwen 3.6 35B-A3B. On the same hardware, the Qwen model produces far more tokens per second per GPU.

• Memory for weights — 66.3 GB in bfloat16 for Nanbeige4.3-33B. The Qwen model fits comfortably on a single card when quantised.

• Context — 524,288 tokens native for Nanbeige4.3-33B. 262,144 for Qwen 3.6 35B-A3B, with a 65,536-token maximum output.

• Modality — text in, text out, English and Chinese, for Nanbeige4.3-33B. Text, image and video in for Qwen 3.6 35B-A3B.

• Price — no hosted endpoint for Nanbeige4.3-33B. Qwen 3.6 35B-A3B on OrcaRouter at $0.25 per million input and $1.49 per million output tokens, with a p50 time-to-first-token of 6.67 seconds.

• Licence — Apache 2.0 on both.

Read the compute line twice. A 33B dense model running a 65,536-token agentic response — Nanbeige's own recommended ceiling — does 33 billion parameters' worth of work for every one of those tokens. The Qwen model does roughly a tenth of that. If the two score within a few points of each other on the same task, the sparse model wins on economics by an order of magnitude, and no benchmark table will tell you so.

The scores cannot be subtracted, and the reason is methodological

Nanbeige4.3-33B is at 20 on the Artificial Analysis Intelligence Index v5.0.1. Qwen 3.6 35B-A3B has no score on the v5 composite at all — its model page reads N/A, because the current index only scores configurations with a publicly available API or a completed in-house run, and the last published Artificial Analysis figures for it sit on the older v4.3.x index, in the high teens for the reasoning configuration.

Those numbers are not a comparison. The v5 index is a rebased evaluation set, and a v4.3.x score does not transfer. Anyone presenting "20 against 18" is subtracting across two different exams, which is the single most common error in model comparison writing and worth naming explicitly here.

What does exist on shared ground is narrower and still useful.

• Terminal-Bench — Artificial Analysis ran Qwen 3.6 35B-A3B at 41.6% on Terminal-Bench 2.1 and 25.8% on Terminal-Bench Hard. It ran Nanbeige4.3-33B at 29.8% on Terminal-Bench 4.0. Different benchmarks, so still not a delta, but both are outside measurements on the same family of task.

• Vendor Terminal-Bench claim — Nanbeige reports 80.9 on Terminal-Bench 2.1 for Nanbeige4.3-33B, which is the same benchmark Qwen 3.6 35B-A3B was independently measured at 41.6% on. If that vendor row holds up, it is the largest gap in this entire comparison. It has not been independently reproduced, and Nanbeige's own footnote says it was run with a Terminus 2 harness.

• Knowledge — Artificial Analysis scores Nanbeige4.3-33B at 11.7% accuracy on AA-Omniscience with 81.8% non-hallucination. Qwen 3.6 35B-A3B's Omniscience figures predate the v5 rebuild, so they sit on a different scale and cannot be lined up against Nanbeige's. What can be said is directional and matches both families' users: the Qwen model knows more, the Nanbeige model declines to answer more often.

• Reasoning — GPQA Diamond 84.8 with reasoning on and 84.1 at the evaluated setting for Qwen 3.6 35B-A3B. Nanbeige4.3-33B's card claims 89.5, unreproduced.

Why Nanbeige picked Qwen3.8-27B instead

There is a defensible reason to benchmark a 33B dense model against a 27B dense model and not against a 35B mixture of experts, and it is not dishonesty. It is that like-for-like on architecture makes the comparison about training quality rather than about serving efficiency. Two dense models at similar scale tell you which lab trained better. A dense-versus-MoE comparison tells you which design is cheaper, which is a different question with a different answer.

But the choice has a cost for the reader, because most people choosing a model are choosing on total cost of ownership, not on training quality in the abstract. On that axis Qwen 3.6 35B-A3B has an advantage Nanbeige's table never surfaces: 3B active parameters means you serve it on hardware you probably already have, and it accepts screenshots and video frames, which the dense 33B cannot do at any price.

The counter-argument is equally concrete. Nanbeige4.3-33B holds 524,288 tokens natively with no scaling, and a dense stack makes its long-context behaviour predictable in a way a sparse model's is not. At 208 KiB of KV cache per token across 52 layers, its full window costs about 104 GiB — which is a warning, not a selling point, but it also means there is no hidden approximation between what the card promises and what the model does.

Where the decision actually lands

If your work is multimodal, Qwen 3.6 35B-A3B is the answer before any benchmark is consulted. Nanbeige4.3-33B is text-only, and a 3B-active model that reads images at a quarter the price per token is not a close second for a screenshot pipeline.

If your work is long-horizon terminal automation and you have four GPUs, Nanbeige4.3-33B's 29.8% on Terminal-Bench 4.0 is the strongest published outside number in this pair, and the model's 81.8% non-hallucination rate is the trait that keeps an agent loop from confidently doing the wrong thing for twenty minutes. That is a real advantage and it is not cheap.

For everyone in between, the practical answer is that Qwen 3.6 35B-A3B is the model you can evaluate this afternoon and Nanbeige4.3-33B is the model you evaluate after a hardware purchase. Qwen 3.6 35B-A3B is in the OrcaRouter catalogue at the price above, on the same key as the rest of the 200-plus models, with automatic failover across upstream providers and a routing DSL if you want to split traffic between it and a second model without writing a second client. Nanbeige4.3-33B is not in the catalogue — we do not host it, and nothing in this article claims we do.

That asymmetry is the honest summary of the whole comparison. One of these models is a service you can meter and switch away from in an afternoon. The other is a checkpoint you adopt. Nanbeige built the second kind deliberately: it is a teacher, its card says so, and the 3B student that will make it matter has not shipped yet.