Nanbeige ships a 33B teacher under Apache 2.0, and its 512K window costs 104 GiB of cache
Nanbeige LLM Lab put a 33B dense agent model on Hugging Face under Apache 2.0 with a 512K window, a self-scored twenty-row card, and a 104 GiB cache bill.
Vlad Makarovreviewed and published
The weights went up on October 24 and the card was finished the next day: a 33-billion-parameter dense model from Nanbeige LLM Lab, Apache 2.0, a 524,288-token window, thinking on by default. By any reading of the model card, this is the most substantial thing the lab has published, and it is aimed squarely at agent work rather than chat.
The card carries a twenty-row scorecard against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, and Nanbeige's own score is bolded in seventeen of those rows. What the card does not carry is a number that follows from its own configuration: at 512K tokens the key-value cache for this architecture is about 104 GiB, on top of 66.3 GB of weights.
What landed in the repository
Fourteen safetensors shards totalling 66.3 GB, stored in bf16. The config declares LlamaForCausalLM: 52 layers, hidden size 7,168, 56 attention heads over 8 key-value heads at dimension 128, an FFN of 22,016, a vocabulary of 166,912 and untied embeddings. Rotary base is 120,000,000 with no rope_scaling block, which is the usual way a vendor claims a window was trained rather than stretched. A separate base checkpoint sits alongside the chat model.
The lineage is worth stating plainly. Every earlier public checkpoint in the organisation is small — 3B models plus a 4B verifier — and Nanbeige4.2-3B alone has 116,877 downloads. The card describes the 33B as the "flagship teacher" that will anchor distillation into a coming Nanbeige4.3-3B. The lab is the model arm of the recruitment platform BOSS Zhipin; the contact address on the card is a kanzhun.com mailbox. A 33B teacher is a supporting asset for that plan, not the product.
The scorecard is the lab's own
Every figure below is from the vendor's table, printed as the card prints it, with the harness the vendor names in its footnotes. Terminal-Bench 2.1 was run with Terminus 2. Terminal-Bench 4 and DeepSWE used mini-swe-agent v2.4.6. NL2Repo ran inside Claude Code. SWE-Atlas-QnA answers were graded against 1,304 rubrics by GPT-6 Sol at medium reasoning effort. No independent reproduction is attached.
| Benchmark | Nanbeige4.3-33B | Qwen3.8-27B |
|---|---|---|
| Terminal-Bench 2.1 | 80.9 | 73.0 |
| Terminal-Bench 4 | 23.8 | 6.3 |
| DeepSWE | 65.5 | 42.5 |
| SWE-Bench Pro | 60.3 | 61.7 |
| GDPval | 84.8 | 86.4 |
| GPQA Diamond | 89.5 | 91.3 |
| AA-LCR | 84.0 | 82.0 |
Read the harness list before the scores. Automation Bench was run on 600 released tasks, Claw-Eval only on its 157-task General subset, GDPval on revised rubrics, and YC-bench kept just the ten most recent user turns in context. Each choice is defensible on its own. Together they mean these are controlled internal comparisons, not positions on a public leaderboard, and the card says as much by footnoting every adapted row.
Where the lead narrows
The wins are real and unevenly distributed. On Terminal-Bench 4 the model scores 23.8 against 6.3, which is a wide relative margin and a failing grade in absolute terms: it does not finish most of the tasks. DeepSWE shows 65.5 against 42.5, the largest gap in the coding group.
Then the margins collapse. SWE-Bench Pro puts Qwen ahead at 61.7 to 60.3, GDPval has Qwen at 86.4 to 84.8, Tau3-Bench is 48.8 to 48.5 and EnterpriseOps-Gym 41.0 to 40.2. Gemma4-31B takes GPQA Diamond at 91.3 to 89.5, and AIME 2026 is 95.2 to 94.5 against Qwen, which is noise at that ceiling. The honest summary of the table is that Nanbeige bought a lead on long-horizon terminal and repository tasks and arrived level everywhere else.
512K is a memory claim before it is a capability claim
The cache arithmetic is not in dispute, because the config supplies every term. Eight key-value heads of dimension 128 across 52 layers, in bf16, is 208 KiB per token. A full 524,288-token sequence therefore needs roughly 104 GiB of cache — about 1.7 times the weights — and 26 GiB at 128K, 6.5 GiB at 32K.
That is why the card's own serving recipes for SGLang and vLLM set tensor parallelism to four. The window is a datacentre feature, priced per token. The card lists one long-context benchmark, AA-LCR at 84.0, and says nothing about what retrieval quality survives at the far end of the window.
Two details make the window load-bearing rather than decorative. Every benchmark number was produced with preserve_thinking enabled, so the model's earlier reasoning stayed in context across turns, which is precisely the setting that turns a long window into a cache bill. And the card publishes no throughput figure at all, so what a four-turn agent run inside that context costs in wall-clock time is left to whoever tries it.
The local story is thinner than the download counts
Quantized copies appeared within a day, and they are the ones moving. AtomicChat's GGUF repository shows 8,885 downloads against 2,787 for the original bf16 repository, with a further abliterated build and an NVFP4 conversion adding more.
The quantizer's own notes undercut the enthusiasm. It reports the model as unusually sensitive to quantization: llama.cpp's plain Q6_K reproduces the original's next token only 95.5% of the time, where most models reach 98% at six bits, and plain Q4_K_M manages 89.9%. Its calibrated IQ4_XS file is 20.3 GB and still changes the next token 8% of the time. That file is 18.9 GiB, so a 24 GiB card has roughly 5 GiB left for cache: about twenty thousand tokens, or under 4% of the advertised window.
What the card does not say
No training token count, no data mixture, no post-training or reinforcement-learning recipe, no throughput or latency figures, no safety evaluation, no technical report. Apache 2.0 on the tensors is genuine and commercially usable; the process that produced them is closed, which makes this open-weight in the strict sense and not open source.
None of that is disqualifying, and the licence is more permissive than most of what shipped this year. It does set the terms on which the numbers should be read: a self-reported twenty-row table, on adapted harnesses, from a lab whose previous ceiling was 4B, with no throughput data and no independent rerun. The first useful test is not another vendor table. It is somebody outside the lab reproducing 23.8 on Terminal-Bench 4 and 65.5 on DeepSWE, then reporting what the cache bill looked like at 128K.
