Nanbeige4.3-33B vs Xing 4.0 29B-A4B: The 23-Point Terminal Gap and the 16-Point Reversal

Nanbeige4.3-33B and Xing 4.0 29B-A4B are the same pitch delivered from opposite directions. The first is a 33-billion-parameter dense model from BOSS Zhipin's Nanbeige LLM Lab, published on 24 October 2026 under Apache 2.0, with a 524,288-token context and a serving recipe that assumes four GPUs. The second is a 29-billion-parameter mixture of experts from China Telecom's TeleAI, weights public since 16 September 2026, that activates roughly 4 billion parameters per token and whose own card claims it fits a single 24GB consumer card at 4-bit. Both target agentic coding and long tool loops. Both ship open weights. Neither has been put head to head by anyone, including their own labs — Nanbeige's card benchmarks against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B; Xing's benchmarks against Gemma4-26B-A4B and Qwen3.6-35B-A3B. They never mention each other.

That is not a gap you can paper over with a comparison table, because the two vendor tables use different harnesses, different sampling and different competitor sets. What you can do is narrower and more useful: find the rows the two labs happen to share, read the footnotes that decide whether those rows mean anything, and then separate the one model that has been measured by somebody who was not paid to measure it from the one that has not.

33 billion dense against 4 billion active, before any benchmark runs

The architectural difference is not a talking point. It is the bill, the latency and the deployment plan, and it is decided before either model answers a single prompt.

Nanbeige4.3-33B is a standard dense decoder — LlamaForCausalLM, 52 layers, hidden size 7,168, grouped-query attention at 56 query and 8 KV heads. Every generated token pays for all 33.1 billion parameters. The BF16 checkpoint is 66.3 GB across fourteen safetensors shards, and both the vLLM and SGLang recipes in the repository specify tensor parallelism of 4, which tells you the intended machine is a four-GPU node rather than a workstation.

Xing 4.0 29B-A4B is sparse: 29 billion parameters in the checkpoint, 64 routed experts with 4 active plus one shared, 40 layers, hidden size 3,584, built on what its card calls mHC + MLA + MTP. Per-token compute scales with the 4 billion active parameters, not the 29 billion stored ones. The vendor claims a 4-bit quantisation at roughly 15 GB of VRAM, which would put it on an RTX 3090 or 4090 — a claim worth checking against the actual quantised file before you buy hardware around it, because the repository reports a 31B-parameter tensor size for the full checkpoint and the two numbers sit oddly together.

Everything else follows from that split.

• Context — Nanbeige4.3-33B 524,288 tokens, native, with no YaRN extension in the recipe; Xing 4.0 29B-A4B 262,144 native, described as extensible to 512K, though no extension recipe is published.

• Minimum practical footprint — Nanbeige4.3-33B 66.3 GB of BF16 weights before KV cache, four-way tensor parallel; Xing 4.0 29B-A4B roughly 15 GB at 4-bit on the vendor's own claim, one card.

• Training stack — Nanbeige4.3-33B is not documented in detail; Xing 4.0 29B-A4B was trained end to end on Ascend 910C clusters with MindSpore and MindFormers, with a claimed 96% throughput improvement over the vendor's own out-of-the-box baseline. That is an internal-baseline claim with no public reference point, so it is unverifiable from outside, but it is the more interesting engineering statement of the two releases.

• Licence — Apache 2.0 on both, which makes the download price zero for both and moves the argument entirely to hardware and tokens.

Five rows both labs ran, and the two that disagree

The two model cards share five benchmark names: Terminal-Bench 2.1, Tau3-Bench, Claw-Eval, AIME2026 and AA.LCR. Reading across them is the only quantitative comparison available, and the footnotes are where it either holds up or falls apart.

• Terminal-Bench 2.1 — Nanbeige4.3-33B 80.9, Xing 4.0 29B-A4B 57.5. Both labs say they ran it in Terminus 2, which makes this the closest thing to a shared instrument in either table. It is still not the same experiment: Xing's card specifies temperature 0.8, top-p 0.95, max 64K tokens, a 24-hour timeout and the average of three runs; Nanbeige's names the harness and nothing else. A 23-point gap under near-identical harnesses is the strongest single signal in this comparison, and it is still a signal rather than a measurement.

• Tau3-Bench — Nanbeige4.3-33B 48.8, Xing 4.0 29B-A4B 64.63. This is the reversal, and it is large. Xing wins tool-use dialogue by nearly 16 points, on the benchmark family that most directly tests multi-turn function calling. The caveat is the metric: Xing reports pass^1 averaged over four runs on the official tau-bench harness; Nanbeige says it used the official repositories but does not state the aggregation, and lists the row under agentic coding rather than tool use.

• Claw-Eval — Nanbeige4.3-33B 78.3 at pass^3, Xing 4.0 29B-A4B 76.55. This one should be discarded rather than read. Nanbeige's footnote says it evaluated the Claw-Eval General subset only, 157 single-turn tasks; Xing's does not name a subset. A 1.75-point difference between two unnamed versions of the same benchmark is noise wearing a result's clothing.

• AIME2026 — Nanbeige4.3-33B 95.2, Xing 4.0 29B-A4B 90.0. Competition mathematics, five runs averaged on the Xing side, max 131,072 tokens; Nanbeige's sampling is again unstated. Both are strong, the ordering is consistent with the terminal result, and neither number tells you much about either model's actual job.

• AA.LCR — Nanbeige4.3-33B 84.0, Xing 4.0 29B-A4B 61.0. Long-context reasoning, and the widest proportional gap after Terminal-Bench. Worth flagging that this is Artificial Analysis's benchmark being run by two vendors, so neither figure is an Artificial Analysis figure.

Four of the five favour Nanbeige4.3-33B. The one that reverses is the one about calling tools in sequence, which is what an agent does all day. That contradiction is the most useful thing on either card, and neither lab has addressed it.

There is a credibility note that cuts in Nanbeige's favour on the terminal row specifically. Its own card claims Terminal-Bench 4.0 at 23.8; Artificial Analysis, running its own harness, measured 29.8. Outside measurement came in six points higher than the vendor's own claim, which is the opposite of the usual direction and suggests the vendor table here is not inflated. Xing 4.0 29B-A4B has no equivalent check.

Only one of these models has a number nobody paid for

Nanbeige4.3-33B is scored. Artificial Analysis put it at 20 on the Intelligence Index v5.0.1 — first of nine in its open-weight comparison class against a class median of 10, and 76th on the all-models list. The sub-scores are more informative than the composite: Terminal-Bench 4.0 at 29.8%, FrontierCode 2.0 at 10.7%, and a non-hallucination rate of 81.8%, which is the seventh-best figure on the site. It declines a lot of questions. For a model explicitly built as a distillation teacher, that is a feature.

Xing 4.0 29B-A4B has no Artificial Analysis entry. Not a poor one — none. The site returns 404 for the model's slug in both plausible spellings, and no composite, sub-score, price or speed figure exists for it anywhere on that board.

That asymmetry governs everything a comparison article can honestly claim. The two models can be compared on vendor tables, with the footnote caveats above. They cannot be compared on the one instrument that neither lab controls, because only one of them is on it. Any piece that ranks these two against each other on neutral ground is inventing the neutral ground.

Where the bill actually diverges

If you are choosing between them, the price of a token is not the deciding number — the cost of finishing a task is, and three separate multipliers stack on the Nanbeige side.

The first is density: 33 billion parameters read per output token against roughly 4 billion. Whatever hardware you own, that is a large multiple in compute per token before anything else is considered.

The second is verbosity, and it is measured rather than assumed. Artificial Analysis reports that Nanbeige4.3-33B generated 160 million output tokens across its Intelligence Index run against a class median of 77.7 million — 2.06 times the median, and eighth of eight on the site's verbosity ranking. The model's own card explains part of it: thinking is on by default, and the recommended ceiling is 262,144 tokens for reasoning work. Any model told it may think for a quarter of a million tokens will use some of them.

Put those two together and a dense model that also talks twice as much is not twice as expensive to run per completed task; it is closer to an order of magnitude. That is the real content of the 33B-versus-4B-active comparison, and it does not appear anywhere in either benchmark table.

The third is unknown, and it is unknown in Xing's favour only by default. No one has published a verbosity figure for Xing 4.0 29B-A4B, so a 4B-active model with no token-efficiency measurement could be genuinely cheap per task or could emit enough tokens to erase part of the advantage. You are comparing a measured expensive model against an unmeasured cheap one.

What you can call today

Neither model is hosted in the ordinary sense, and the difference between them is smaller than the launch coverage suggests.

Nanbeige4.3-33B has no API at all. Its Hugging Face page says the model is not deployed by any inference provider, and Artificial Analysis records no publicly available API and leaves the price fields blank. Access means downloading 66 GB of weights and running them on hardware you own or rent.

Xing 4.0 29B-A4B has weights since 16 September and a metered route that appeared in mid-October: a third-party marketplace lists it at $0.22 per million input tokens and $0.90 per million output, and a free promotional variant of the same endpoint has been listed alongside it. The Hugging Face inference-provider panel is still empty, so the hosted path runs through marketplaces rather than the vendor, and a promotional tier is not a price you should build a budget on.

To be direct about our own position: neither model is in OrcaRouter's catalogue today, so we are not a route to either one, and nothing below pretends otherwise. Where a router earns its place in a decision like this one is as the control. The cheapest way to find out whether a self-hosted 33B dense model is worth four GPUs is to put it beside something you can call per token and measure both on your own tasks — and the obvious control here is the model Nanbeige's own card benchmarks against. Qwen3.8-27B is in our catalogue at $0.33 per million input and $2.40 per million output, Artificial Analysis scores it 20.1 against Nanbeige's 20, and it takes the same key as the rest of the catalogue with provider list prices passed through at no markup. That makes the comparison a routing rule rather than a second contract.

Who should pick which

The split is legible once you accept that this is a capability-versus-cost decision rather than a capability-versus-capability one.

Nanbeige4.3-33B is the pick when terminal work is the product. A 23-point lead on a near-shared Terminus 2 harness, the highest non-hallucination figure in its class, a native 512K window and an independent score that came in above the vendor's own claim — that is a model worth four GPUs if your workload is long-horizon code and shell work, and if you can absorb a token bill roughly twice the class median.

Xing 4.0 29B-A4B is the pick when a single card and a metered invoice are the constraints. A 16-point lead on Tau3-Bench, a footprint the vendor says fits a 24GB card, a quarter of Nanbeige's per-token compute, and the only one of the two with a price you can pay without buying hardware. It is also the weaker model on the one row where the harnesses most nearly match, so the case for it is economic and operational, not capability.

Two things would settle the parts that remain open. An Artificial Analysis run on Xing 4.0 29B-A4B would turn an asymmetric comparison into a symmetric one, and a Tau3-Bench number from Nanbeige with a stated aggregation would tell you whether the 16-point reversal is a real weakness in tool use or an artefact of two different metrics sharing a name. Until either lands, the honest summary is that Nanbeige4.3-33B is the stronger model on the evidence that exists and Xing 4.0 29B-A4B is the one you can afford to run — and the size of that trade is a question about your task mix, not about either card.