Nanbeige LLM Lab put a 33B dense model on Hugging Face under Apache 2.0, with 512K of context and a benchmark table that wins 17 of 20 tests. The same release calls it the teacher for a 3B model coming later.
Nanbeige4.3-33B Lands Under Apache 2.0
Nanbeige4.3-33B went up on Hugging Face on October 25 with an Apache 2.0 license, alongside a Base variant meant for custom fine-tuning. The weights are open, the architecture is standard, and the license allows commercial use.
The interesting part isn't the license. Plenty of open-weight models ship Apache 2.0. What's unusual is the spec sheet: a 33B dense model with a 524,288-token context window, positioned squarely at agentic coding rather than chat.
Who Built Nanbeige4.3-33B
Nanbeige LLM Lab sits inside BOSS Zhipin, the Chinese recruiting platform. That's a job-matching company training a coding agent, which sounds odd until you remember what recruiters are starting to automate: screening, structured extraction, and interviews that need software to reason through a candidate's work.
The lab has shipped earlier Nanbeige models, and this one continues the naming without much ceremony. There's no launch blog post, no demo site. A model card, a base model card, and a post on X.
For a western reader, that's a signal about distribution. This model is aimed at developers who read model cards, not at app users.
Why a Dense 33B Model Costs More to Run Than an MoE
Dense means every parameter runs for every token. A mixture-of-experts model of the same total size wakes up a fraction of its parameters per token, so it answers faster and cheaper at inference while carrying the same weights on disk.
Which is why most labs in 2026 ship sparse models at this size and call it a day. A 33B dense model needs memory and compute for all 33 billion weights on every request.
Nanbeige chose dense anyway, and the model card doesn't explain why in detail. The usual reason is training simplicity and cleaner distillation targets: a dense teacher produces one consistent output distribution instead of routing between experts.
The cost lands on whoever deploys it. You'll pay per token in a way a 33B MoE wouldn't charge you.
Nanbeige4.3-33B Benchmarks: 17 Wins and Three Losses
All scores below are self-reported by Nanbeige, measured against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at their highest reasoning settings, and none have been independently reproduced.
Benchmark | Nanbeige4.3-33B | Qwen3.8-27B |
|---|---|---|
Terminal-Bench 2.1 | 80.9 | 73.0 |
DeepSWE | 65.5 | 42.5 |
Automation Bench | 54.2 | 38.3 |
JobBench | 44.8 | 34.5 |
SWE-Bench Pro | 60.3 | 61.7 |
Terminal-Bench 4.0 | 23.8 | not listed |
The wins are large where the model is aimed. Terminal-Bench 2.1 tests command-line work across many steps, and an eight-point lead there says something about holding a plan together. DeepSWE at 65.5 against 42.5 is a gap that would be hard to believe if it holds up under a rerun.
The losses matter just as much. SWE-Bench Pro, which measures real repository fixes, goes the other way, 60.3 to 61.7. GDPval is close but behind, 84.8 against 86.4. And Terminal-Bench 4.0 at 23.8 is the lowest number in Nanbeige's own table, which suggests the hardest tier of tool use is still rough.
A 33B model beating a 27B on agent tasks and losing on patch quality is a believable profile. It's the difference between planning well and landing the edit.
Two Rows in the Nanbeige Table Look Swapped
There's a problem in the comparison table that no amount of difference in test setup explains.
Nanbeige lists Meta's Muse-Glimmer-30B at 51.2 on SWE-Bench Pro and Gemma4-31B at 36.9. Meta's own numbers put those scores the other way around. The table also gives Gemma4-31B a 91.3 on GPQA Diamond, where Google reports 84.3.
Different test setups can shift a score by a few points. They don't swap two entries.
Two rows, reversed.
Until someone reruns these benchmarks across all four models on a single rig, treat the whole table as unverified, including the wins.
That's not a reason to dismiss the release. It's a reason to run your own eval before you swap a production model.
512K Context, 66 GB of Weights and a Cache Problem
The context window is the headline number: 524,288 tokens. The BF16 weights are about 66 GB, which fits on a single 80 GB card with room to spare, and that detail is probably why people assume the whole thing runs comfortably on one GPU.
The cache is where it gets painful. Nanbeige's own math works out to roughly 208 KiB per token of KV cache at full context, which is about 112 GB at 512K. Suddenly the weights are the small part of the bill.
In practice that means no single consumer card runs this at full context, and multi-GPU serving becomes the default for anything long, which is why the model card's deployment section reads like a data center recipe rather than a hobbyist guide. The model card ships recipes for Transformers, SGLang and vLLM, with a four-way tensor parallel example.
There is a middle path. Community Q8 quantizations land around 35.2 GB, and with a 32K context that runs on one 48 GB card. For agent work with bounded context, that's the configuration most people will actually use.
Architecture details, for the people who care: 52 layers, hidden size 7,168, 56 query heads against 8 KV heads in grouped-query attention, a 166,912-token vocabulary, and RoPE theta at 1.2e8. Standard LlamaForCausalLM, so existing tooling works without patching.
At release there was no inference provider hosting it, which means you're downloading 66 GB or you're not trying it.
The Teacher Model Plan: 33B Now, 3B Later
The model card frames this release as step one. Train a strong 33B agent, then run teacher-student distillation into Nanbeige4.3-3B, which hasn't shipped yet.
Distillation is the standard trick for putting agent behavior on small hardware, and it works better when the teacher is dense. That explains the architecture choice more convincingly than anything else in the release, since a dense teacher hands a student one clean target instead of a routing pattern that has to be learned twice.
For now, the 3B model is a promise. Nothing to download, nothing to test.
Who Should Download Nanbeige4.3-33B
If you run coding agents on self-hosted hardware and you have 80 GB of GPU memory, this is worth an afternoon. The Terminal-Bench numbers point at exactly the workload you'd use it for, and Apache 2.0 means no legal review.
If you're on a laptop, wait for the 3B. If you need long-context retrieval at full 512K, budget for the cache before you budget for the GPU.
What would you actually test first? For most teams it's the patching loop: ask the model to fix a bug in a repo it hasn't seen, then read the diff. A model that plans well and patches badly will look great on Terminal-Bench and useless in your IDE, which is exactly what the SWE-Bench Pro gap hints at.
The model you can inspect is worth more than the leaderboard you can't. That's the case for downloading it, and the case for doubting the table it came with.
