Nanbeige4.3-33B vs Nanbeige 4.2-3B: The 3B Wins One Benchmark and the 33B Explains Why It Exists

There is a benchmark in Nanbeige's own published tables where the small model beats the large one, and it is worth starting there because it tells you what kind of release Nanbeige4.3-33B actually is. Nanbeige 4.2-3B scores 63.6 on SWE-Bench Verified. Nanbeige4.3-33B does not report SWE-Bench Verified at all. It reports SWE-Bench Pro, where it scores 60.3 — and Nanbeige 4.2-3B's SWE-Bench Pro figure is 46.9. So the comparison is 60.3 against 46.9 on the harder variant, and the 3B's impressive 63.6 is answering a different question.

That is the shape of the relationship between these two models. Nanbeige4.3-33B is not a bigger Nanbeige 4.2-3B. It is a different architecture, a different benchmark suite, a different licence story and — most importantly — a different purpose. The 3B is a product you can run. The 33B is a teacher the lab built so it can distil the next 3B, and the card says so in its second paragraph.

The architectures are not the same family of design

Nanbeige 4.2-3B is a Looped Transformer: a single 22-layer stack executed twice, giving 44 effective layer passes and 44 KV slots from 22 layers of weights. That is how it reaches 4B total parameters and 3B non-embedding parameters while behaving like something larger. It has 262,144 tokens of context, Apache 2.0, and was trained on 28 trillion tokens with math, code, STEM and synthetic QA upsampled.

Nanbeige4.3-33B is a plain dense Llama stack. Fifty-two layers, hidden size 7,168, 56 query heads against 8 key-value heads, an FFN width of 22,016, a 166,912-token vocabulary, RoPE base 120 million with no scaling, and 524,288 positions. It is 33.1 billion parameters, 66.3 GB in bfloat16, and there is no looping anywhere in it.

So the lab moved away from the trick that defined its small models. That is not a small editorial decision. Looping buys capacity per parameter and costs memory per token of context; a dense stack buys predictability and costs parameters. For a teacher model whose job is to emit millions of high-quality trajectories, predictable attention and a huge window are worth more than parameter efficiency, because the teacher is not the thing that has to fit on a laptop.

• Parameters — 33.1B dense for Nanbeige4.3-33B vs 4B total and 3B non-embedding for Nanbeige 4.2-3B.

• Architecture — 52-layer conventional attention for Nanbeige4.3-33B vs a 22-layer stack run twice for Nanbeige 4.2-3B.

• Context — 524,288 tokens for Nanbeige4.3-33B vs 262,144 for Nanbeige 4.2-3B.

• Weights on disk — 66.3 GB bfloat16 across fourteen shards for Nanbeige4.3-33B vs a few gigabytes for Nanbeige 4.2-3B.

• Licence — Apache 2.0 on both, so nothing changes for a commercial user moving between them.

• Training disclosure — Nanbeige4.3-33B publishes no pretraining token count, no data mix and no compute figure. Nanbeige 4.2-3B publishes 28 trillion tokens, a paper (arXiv 2607.22083) and a full evaluation methodology.

That last row is the one that should give a reader pause. The newer, larger, more capable model is documented far less thoroughly than the older, smaller one. The 4.2-3B release came with a technical report; the 4.3-33B release came with a model card and nothing else.

What the parameter gap actually buys

Comparing the two models on the benchmarks they both report is the fairest exercise available, and the results are more nuanced than an eleven-fold parameter difference would suggest.

• GPQA Diamond — Nanbeige4.3-33B 89.5 vs Nanbeige 4.2-3B 87.4. A 2.1-point gap for 8× the parameters.

• GDPval — Nanbeige4.3-33B 84.8 on the vendor table vs Nanbeige 4.2-3B 74.3 on the rubrics variant. Around 10 points, though the two cards do not describe the rubric set identically.

• AA-LCR — Nanbeige4.3-33B 84.0 vs Nanbeige 4.2-3B 58.7. This is the widest gap in the comparison, and it is the long-context one.

• SciCode — Nanbeige4.3-33B 52.6 vs Nanbeige 4.2-3B 35.6. A 17-point gap on scientific coding.

• LiveCodeBench v6 — Nanbeige4.3-33B does not report it. Nanbeige 4.2-3B scores 72.5.

• HLE without search — Nanbeige4.3-33B does not report it. Nanbeige 4.2-3B scores 17.8.

The pattern is that the 33B's advantage concentrates in long context and long-horizon code, and nearly vanishes in knowledge-heavy multiple choice. GPQA at 87.4 for a 3B-class model is an extraordinary number, and it is the clearest evidence that the looped design works. AA-LCR at 58.7 against 84.0 is equally clear evidence that looping costs you when the context gets long, which is exactly the trade the architecture makes.

Where the 33B separates itself decisively is on the newer, agentic benchmarks that only it reports: Terminal-Bench 2.1 at 80.9, Terminal-Bench 4.0 at 23.8, DeepSWE at 65.5, JobBench at 44.8, AutomationBench at 54.2, Toolathlon-Verified at 70.4. Nanbeige 4.2-3B has its own agentic rows — Claw-Gym 65.0, MCP-Atlas 57.8, Claw-Eval pass^3 52.2, Pinch-Bench-V2 74.7 — and they are respectable for a 3B. They are not the same tasks, and no one has run both models through one harness.

The independent data covers only one of them

Artificial Analysis has scored Nanbeige4.3-33B at 20 on the Intelligence Index v5.0.1, with detailed sub-scores: Terminal-Bench 4.0 at 29.8% on its own harness — six points above Nanbeige's own 23.8 claim — FrontierCode 2.0 at 10.7%, AA-Omniscience at −4.4 net with 11.7% accuracy and 81.8% non-hallucination, and a verbosity figure of 160 million output tokens against a class median of 77.7 million.

Nanbeige 4.2-3B has no entry on the v5 composite. Its card claims the top spot on Artificial Analysis's leaderboard for small models, and that claim appears to refer to an earlier index or a small-model board rather than the v5 Intelligence Index. Treat it as a vendor statement until you can find the row.

So the only model in this pair with a verified outside number is the 33B, and the number is 20 — mid-table overall, first of nine in its comparison class.

Which one you would actually deploy

This is where the comparison stops being academic. Nanbeige 4.2-3B runs on hardware people own. There are 59 quantisations in its model tree, a LiteRT-LM on-device port, day-one recipes for llama.cpp, Ollama and LM Studio, and a third-party Apple Silicon study that patched five deployment bugs in the released checkpoint and measured MCPMark at 3/10 on the easy subset afterwards. That study is worth reading before you ship the 3B: it found a silently zeroed RoPE buffer, removed cache APIs, a chat-template regression that breaks multi-tool calls, and roughly double the attention memory from looping.

Nanbeige4.3-33B runs on four 80 GB GPUs in bfloat16, or one 48 GB card at Q5 with an importance matrix. The quantisation profile is worse than the 3B's: a community GGUF repository reports that the model is unusually sensitive to quantisation, with importance-matrix Q4_K_M at 92.6% top-1 agreement against bfloat16 and plain Q4_K_M at 89.9%. For tool calls, where one malformed argument ends a turn, that difference is the product.

Neither model is in the OrcaRouter catalogue. We do not host Nanbeige4.3-33B and we do not host Nanbeige 4.2-3B, and nothing here should be read as a claim that we do. If you are choosing between them for a self-hosted deployment, the useful move is to put whatever you land on behind an OpenAI-compatible gateway alongside a hosted fallback, so a failed quantised tool call degrades to a working model instead of to an error page. That is the same routing surface the catalogue runs on, and the 0% markup matters there in a specific way: the hosted comparison keeps the vendor's list price, so the cost you measure against your GPU bill stays honest when a price moves.

The model that will actually settle this

Nanbeige4.3-3B is the point of the whole exercise. The 33B's card describes it as the flagship teacher built so the lab can distil agentic capabilities into smaller models, "e.g. for Nanbeige4.3-3B", and the benchmark suite it published — heavy on Terminal-Bench, DeepSWE, JobBench, Toolathlon — is a list of the behaviours someone wants to transfer. When that student ships, it will be the model most people run, and the 33B will be the thing that explains its scores.

Until then the choice between these two is not close for most teams. If you need a model today that runs on a workstation and handles agentic tasks competently, Nanbeige 4.2-3B is available and cheap and has a paper behind it. If you need a 512K window, a dense attention stack and the strongest long-horizon agentic scores the family has published, Nanbeige4.3-33B is the only option and you will pay for it in GPUs. What you should not do is read the 33B's arrival as making the 3B obsolete. The 3B is the shipping product. The 33B is the factory.