Nanbeige releases Nanbeige4.3-33B, an Apache 2.0 dense model for agentic coding with a 512K-token context
Nanbeige LLM Lab, the large language model research team of the Chinese recruitment platform BOSS Zhipin, has published Nanbeige4.3-33B on Hugging Face under the Apache 2.0 license. The weights were uploaded to the repository on October 24, 2026, and the model card with the full benchmark comparison was finalized on October 25. The release is a 33-billion-parameter dense transformer with a context window of up to 524,288 tokens (512K), trained for agentic coding, tool use and what the team calls professional "cowork" tasks, and it supports English and Chinese.
The team describes Nanbeige4.3-33B as the flagship "teacher" of the Nanbeige4.3 series. Its stated purpose is not only to be used directly but to anchor a distillation pipeline for smaller models, with a forthcoming Nanbeige4.3-3B named as the first example. That framing explains a release that departs sharply from the lab's previous public work, which until now consisted almost entirely of 3B-class models such as Nanbeige4-3B, Nanbeige4.1-3B and Nanbeige4.2-3B.
According to the model card, Nanbeige4.3-33B delivers "state-of-the-art agentic performance among open models in the ~30B parameter class." The supporting evidence is a 20-benchmark table in which the model is compared with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all run at their highest thinking effort. Nanbeige4.3-33B posts the best score on most rows, including 80.9 on Terminal-Bench 2.1 and 65.5 on DeepSWE, while Qwen3.8-27B stays ahead on SWE-Bench Pro and GDPval and Gemma4-31B leads on GPQA Diamond. All of these numbers are vendor-reported; no independent evaluation had been published at the time of writing.
··········
A CONVENTIONAL DENSE ARCHITECTURE WITH A VERY LONG CONTEXT.
The published configuration shows that Nanbeige4.3-33B does not use a mixture-of-experts design, which has become the default for many recent open-weight releases. The checkpoint is declared as a `LlamaForCausalLM` model, so every parameter participates in every token. The network has 52 transformer layers with a hidden size of 7,168, 56 attention heads of dimension 128 and a SwiGLU-style feed-forward block with an intermediate size of 22,016. Attention uses grouped-query attention with 8 key-value heads, which means each key-value head is shared by seven query heads and the key-value cache is one seventh of the size it would be with full multi-head attention.
The tokenizer has a vocabulary of 166,912 entries, input and output embeddings are not tied, and the weights are stored in BF16. Positional encoding relies on rotary embeddings with a base frequency (rope_theta) of 120,000,000 and no additional RoPE scaling configured, while both max_position_embeddings and seq_length are set to 524,288. In practice the very high rotary base is how the model reaches its 512K window natively, rather than through an inference-time extension scheme. The model card does not say how many tokens the model was trained on, what data mix was used, or how the long-context stage was organized, and no technical report has been linked.
........
Parameter | Value |
Architecture | Dense decoder (LlamaForCausalLM) |
Parameters | 33B (BF16) |
Layers | 52 |
Hidden size | 7,168 |
Attention heads / KV heads | 56 / 8 (GQA) |
Head dimension | 128 |
FFN intermediate size | 22,016 |
Vocabulary | 166,912 tokens |
RoPE base | 120,000,000 |
Maximum context | 524,288 tokens |
License | Apache 2.0 |
........
The headline size is consistent with the configuration. A Data Studios calculation from the published dimensions gives roughly 590 million parameters per layer, about 30.7 billion across the 52 layers, plus close to 2.4 billion in the two untied embedding matrices, for a total of approximately 33.1 billion. That arithmetic also gives a sense of the memory cost of the long context: with 8 key-value heads of dimension 128 across 52 layers, the BF16 key-value cache needs about 208 KiB per token, which approaches 104 GiB for a single sequence at the full 524,288 tokens, on top of roughly 66 GB of BF16 weights.
··········
BENCHMARK RESULTS ACROSS CODING, COWORK AND REASONING.
The comparison table groups benchmarks into three families: agentic coding, agentic cowork (office automation, planning, enterprise operations and job-like tasks) and classical reasoning. The model card states that all models were evaluated with the same inference hyperparameters and the same harness for each benchmark, and that Nanbeige4.3-33B was run with the preserve_thinking option enabled, so its reasoning from earlier turns remained in context during multi-turn tasks.
........
Benchmark | Nanbeige4.3-33B | Qwen3.8-27B | Muse-Glimmer-30B | Gemma4-31B |
Terminal-Bench 2.1 | 80.9 | 73.0 | 61.8 | 47.2 |
Terminal-Bench 4 | 23.8 | 6.3 | 1.6 | 0.0 |
DeepSWE | 65.5 | 42.5 | 4.4 | 2.7 |
SWE-Bench Pro | 60.3 | 61.7 | 36.9 | 51.2 |
SWE-Atlas-QnA | 50.8 | 41.9 | 17.7 | 3.3 |
NL2Repo | 53.0 | 48.3 | 27.4 | 33.8 |
SciCode | 52.6 | 46.7 | 44.9 | 45.5 |
SkillsBench | 61.1 | 56.4 | 39.7 | 28.9 |
........
On agentic coding, the clearest gaps appear on the hardest terminal and repository tasks. Terminal-Bench 4 is a benchmark where the three comparison models score between 0.0 and 6.3, and Nanbeige4.3-33B reaches 23.8, a low absolute number that still indicates it completes tasks the others almost never finish. On DeepSWE, the 23-point margin over Qwen3.8-27B is the largest in the coding group. SWE-Bench Pro is the exception: Qwen3.8-27B scores 61.7 against 60.3, so on that widely followed software engineering benchmark the two models are effectively level, with a slight edge for Qwen.
........
Benchmark | Nanbeige4.3-33B | Qwen3.8-27B | Muse-Glimmer-30B | Gemma4-31B |
Automation Bench | 54.2 | 38.3 | 15.2 | 9.7 |
JobBench | 44.8 | 34.5 | 16.6 | 9.0 |
GDPval | 84.8 | 86.4 | 75.8 | 71.4 |
Claw-Eval (pass^3) | 78.3 | 75.2 | 57.3 | 44.2 |
Toolathlon-Verified | 70.4 | 68.5 | 47.2 | 27.8 |
DeepPlanning | 63.3 | 49.2 | 21.4 | 19.4 |
EnterpriseOps-Gym | 41.0 | 40.2 | 30.5 | 27.3 |
YC-bench | $1,025,437 | $641,290 | $513,978 | $265,152 |
Tau3-Bench | 48.8 | 48.5 | 33.7 | 18.6 |
GPQA Diamond | 89.5 | 89.2 | 83.5 | 91.3 |
AIME 2026 | 95.2 | 94.5 | 94.7 | 89.2 |
AA-LCR | 84.0 | 82.0 | 83.3 | 72.0 |
........
The cowork results are more uneven in their margins. Nanbeige4.3-33B shows large leads on Automation Bench, JobBench and DeepPlanning, but on Toolathlon-Verified, EnterpriseOps-Gym and Tau3-Bench it is within two points of Qwen3.8-27B, and on GDPval Qwen is ahead by 1.6 points. YC-bench is reported in simulated dollars rather than a percentage, and the Nanbeige figure of about $1.03 million is roughly 60% higher than Qwen3.8-27B's $641,290. The team adds that on JobBench the model outperforms the baseline in most of the 35 occupations evaluated.
On classical reasoning the four models are close. Nanbeige4.3-33B scores 95.2 on AIME 2026 and 84.0 on AA-LCR, a long-context reasoning benchmark, while Gemma4-31B posts the best GPQA Diamond result at 91.3. The pattern suggests that the training budget went primarily into multi-step agent behavior rather than into knowledge-heavy question answering.
··········
HOW THE EVALUATIONS WERE RUN.
The footnotes matter for interpreting the table, because agentic scores depend heavily on the harness. Terminal-Bench 2.1 was run with the Terminus 2 agent, while Terminal-Bench 4, DeepSWE and SWE-Atlas-QnA used mini-swe-agent version 2.4.6; SWE-Atlas-QnA answers were graded against 1,304 rubrics by a judge model at medium reasoning effort. NL2Repo was run inside Claude Code, and JobBench used the Claude Code harness with a search tool added.
Several cowork benchmarks were also adapted. Automation Bench v1.0.6 was evaluated on its 600 publicly released tasks, GDPval was scored with revised rubrics, Claw-Eval was limited to its General subset of 157 single-turn tasks, Toolathlon-Verified was run with no max_tokens limit, and YC-bench kept only the 10 most recent user turns in context. DeepPlanning, EnterpriseOps-Gym and Tau3-Bench used their official repositories. Revised rubrics, subsets and context truncation are reasonable engineering choices, but they mean the numbers are not directly comparable with leaderboard entries produced under different settings, and they should be read as a controlled internal comparison rather than as positions on public leaderboards.
··········
THINKING MODES, TOOL CALLING AND RECOMMENDED SETTINGS.
Nanbeige4.3-33B is a reasoning model with thinking enabled by default. The chat template exposes an enable_thinking switch that can be set to False for a non-thinking mode, and a separate preserve_thinking option that keeps earlier reasoning traces in the conversation. The team recommends leaving preserve_thinking off for ordinary chat and question answering, and turning it on for multi-turn tool use, office tasks and code-agent workflows, which is also the configuration used for the published benchmarks.
Tool calling is enabled by passing tools to the chat template. The recommended format is XML, with JSON also supported. The model card gives two sampling profiles: for agentic and tool-use tasks, temperature 1.0, top_p 0.95 and up to 65,536 new tokens; for reasoning and chat, temperature 0.6, top_p 0.95 and up to 262,144 new tokens. The very large output budgets reflect how long the model's reasoning traces can become on hard problems.
........
Use case | Temperature | top_p | Max new tokens |
Agentic and tool use | 1.0 | 0.95 | 65,536 |
Reasoning and chat | 0.6 | 0.95 | 262,144 |
........
··········
DEPLOYMENT WITH TRANSFORMERS, SGLANG AND VLLM.
The model loads through Hugging Face Transformers with trust_remote_code enabled, automatic dtype and automatic device placement, and prompts are formatted with the tokenizer's chat template. For serving, the model card provides launch commands for SGLang and vLLM that both use tensor parallelism across four GPUs, a 0.8 memory fraction and the Qwen3 reasoning parser together with the qwen3_coder tool-call parser. The following vLLM command is reconstructed from the flags listed in the model card, with the local model path left as a placeholder:
vllm serve <path-to-Nanbeige4.3-33B> \
--host 0.0.0.0 --port 8000 \
--tensor-parallel-size 4 \
--gpu-memory-utilization 0.8 \
--enable-auto-tool-choice \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coderThe reuse of the Qwen3 parsers indicates that the model's reasoning and tool-call output formats are compatible with existing Qwen3 tooling, which lowers the integration cost for teams already running Qwen-family models. The model card also lists Docker Model Runner and community quantizations for llama.cpp, Ollama and LM Studio as options. At the time of capture, no hosted inference provider was serving the model through Hugging Face, so access currently means self-hosting the weights.
The four-GPU recommendation follows from the memory profile. Roughly 66 GB of BF16 weights plus a key-value cache that grows by about 208 KiB per token leaves a single 80 GB accelerator with very little room for long agent sessions, and splitting the model across four devices is the practical way to serve multi-hundred-thousand-token contexts. Quantized community builds will reduce the weight footprint, though their effect on agentic accuracy has not been documented by the team.
··········
A TEACHER MODEL FOR THE NANBEIGE4.3 SERIES.
Nanbeige's previous releases established the lab as a specialist in small models: the Hugging Face organization lists Nanbeige4-3B, Nanbeige4.1-3B and several Nanbeige4.2-3B variants, including FP8 and GPTQ-Int8 quantizations and a DSpark checkpoint published earlier this month. Nanbeige4.3-33B changes the scale of the project, but the team presents it as infrastructure for that same line. The model card states that the 33B model "lays the necessary foundation to power our upcoming small model distillation pipeline," naming Nanbeige4.3-3B as an example, without giving a release date for the smaller model.
The choice of a dense architecture fits that role. A dense teacher has a single, uniform set of weights and produces outputs whose behavior is easier to transfer to a dense 3B student than the routed computation of a sparse model, and it is also simpler to fine-tune and quantize. The trade-off is inference cost: every token activates all 33 billion parameters, whereas mixture-of-experts models of similar total size typically activate only a fraction. The team has not published throughput or latency figures.
The Apache 2.0 license is permissive and allows commercial use, modification and redistribution of the weights. The release is open-weight rather than fully open source in the stricter sense: the training data, data mixture and training code have not been published.
··········
WHAT IS ESTABLISHED AND WHAT REMAINS OPEN.
What can be verified directly is the artifact itself: a 33-billion-parameter dense Llama-architecture model with grouped-query attention, a 166,912-token vocabulary and a native 524,288-token context window, published with BF16 weights under Apache 2.0, with documented thinking controls, XML and JSON tool calling, and serving recipes for Transformers, SGLang and vLLM.
The performance claims are a different category. The 20-benchmark comparison was produced by the developer, uses adapted rubrics, subsets and harnesses for several tasks, and has not yet been reproduced. If the coding results hold under independent testing, particularly on Terminal-Bench 4 and DeepSWE, Nanbeige4.3-33B would be one of the strongest dense open-weight models for agentic software work at this size. The narrow margins on SWE-Bench Pro, GDPval, Toolathlon-Verified and Tau3-Bench suggest that its advantage over Qwen3.8-27B is concentrated in specific long-horizon tasks rather than uniform.
Several important details remain undisclosed: the size and composition of the training corpus, the post-training and reinforcement learning recipe, inference throughput, safety evaluations and the timing of the planned Nanbeige4.3-3B student. Until a technical report or independent evaluations appear, the release is best understood as a well-documented, permissively licensed checkpoint whose headline numbers still require external confirmation.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
