BOSS Zhipin Releases 33B Coding Agent That Claims Top Scores on 17 of 20 Agentic Benchmarks

BOSS Zhipin, the Chinese online recruitment platform listed on Nasdaq, has released Nanbeige4.3-33B — a 33.1-billion-parameter open-weights model designed for agentic coding, terminal-based software engineering, and workplace automation tasks. The model, published October 24 on Hugging Face by the company's internal Nanbeige LLM Lab, claims first place on 17 of 20 benchmarks in a self-reported comparison against Alibaba's Qwen3.8-27B, Google DeepMind's Gemma4-31B, and Meta's Muse-Glimmer-30B. None of those benchmark scores have been independently reproduced.

The release is notable less for its corporate origin — an AI research team embedded inside a job-platform company — than for what its self-reported Terminal-Bench 2.1 score implies. Nanbeige claims 80.9 on that benchmark, which measures real-world CLI task execution across software engineering, data science, and scientific computing scenarios run in a Daytona sandbox. The Terminal-Bench 2.1 public leaderboard shows Claude Opus 4.8 at approximately 78.9 percent on the Claude Code harness — a published frontier reference score. If Nanbeige's self-reported 80.9 were independently reproduced on a comparable harness, it would represent the first 30B-class open-weights model to match or exceed that published frontier score on real terminal task execution, a threshold no open-weights model at that parameter scale has previously cleared.

A Dense Transformer Built for Terminal Agents

Nanbeige4.3-33B uses a standard dense Llama-layout Transformer — not a mixture-of-experts design — with 52 layers, a hidden size of 7,168, and 56 attention heads paired with 8 key-value heads. The grouped-query attention (GQA) configuration compresses KV-cache requirements by a 7:1 ratio, which helps manage memory during long-context inference. The model's vocabulary spans 166,912 tokens, substantially larger than LLaMA-3's 128,256 — a design choice likely intended to improve coverage of both Chinese and English professional terminology.

The context window reaches 524,288 tokens (512K), which gives Nanbeige4.3-33B a significant architectural advantage over the Qwen3.8-27B's 128K window for applications that require processing entire codebases or long document chains in a single pass. Whether attention quality degrades at extreme context lengths — a common weakness in extended-window models — is not addressed in the model card.

The model supports a thinking mode controlled by enable_thinking (default on) and preserve_thinking (recommended true for multi-turn agent sessions), consistent with chain-of-thought inference patterns. Importantly, the lab splits inference parameters by task type: agentic tasks run at temperature 1.0 and a maximum of 65,536 new tokens per turn, while reasoning and chat tasks use temperature 0.6 with a 262,144-token output budget. The higher temperature for agentic tasks suggests the lab found diversity-encouraging sampling beneficial for exploratory terminal execution. Tool calling supports both XML format (recommended) and JSON, making it compatible with common agentic coding frameworks that use either convention. Full inference parameters and tool-use configuration are detailed in the model card.

What the 20-Benchmark Table Actually Shows

The model card presents a 20-benchmark comparison table organized into three categories: Agentic Coding, Agentic Cowork (workplace automation), and Reasoning. Nanbeige claims wins on 17 of the 20 rows — not 18, despite some secondary coverage using that figure. The three benchmarks where Nanbeige does not lead are SWE-Bench Pro (Qwen3.8-27B scores 61.7 versus Nanbeige's 60.3), GDPval (Qwen leads 86.4 versus 84.8), and GPQA Diamond (Gemma4-31B leads 91.3 versus 89.5). On all three, the margin is narrow.

The largest claimed advantage is on DeepSWE, an agentic software engineering benchmark evaluated using mini-swe-agent v2.4.6. Nanbeige self-reports 65.5 versus Qwen's 42.5 — a 54 percent relative improvement. This is the most striking number in the table and also the one requiring most caution, since DeepSWE evaluation methodology varies by agent configuration and the evaluation was conducted entirely by Nanbeige LLM Lab rather than benchmark organizers.

Several benchmarks in the "Agentic Cowork" category are internally developed or narrowly adopted — including Automation Bench, JobBench, and YC-bench (where performance is expressed as a dollar value of outcomes achieved, with Nanbeige reporting $1,025,437 against Qwen's $641,290). The practical meaning of these scores for developers outside BOSS Zhipin's ecosystem is difficult to assess without independent methodology documentation. The Toolathlon-Verified, DeepPlanning, and EnterpriseOps-Gym benchmarks are similarly not yet established as widely accepted third-party evaluations. The full benchmark comparison table is available in the model card.

Graduate-level reasoning results are more straightforwardly interpretable. Nanbeige's GPQA Diamond score of 89.5 places it close to Gemma4-31B's 91.3 and well above Muse-Glimmer-30B's 83.5. On AIME 2026 math competition problems, Nanbeige self-reports 95.2 versus Qwen's 94.5 and Muse-Glimmer's 94.7 — differences that fall within the noise of benchmark evaluation at this score level.

Read more: Moonshot's Kimi K3.1 Pro tops Terminal-Bench among open models with 33% price cut

BOSS Zhipin's Strategic Position as a Non-Traditional AI Lab

Kanzhun Limited, which operates BOSS Zhipin, is not a foundation-model company by origin. Founded in 2013 as a mobile-first recruitment platform, it went public on Nasdaq in June 2021 (ticker: BZ) and now carries a market capitalization of approximately $7 billion. Its core product uses AI matching to connect job seekers and employers — a business generating substantial workplace-workflow data.

The Nanbeige4.3-33B model's "Agentic Cowork" benchmark category and the emphasis on professional task automation suggest BOSS Zhipin is applying that workforce-interaction context to model development. JobBench and EnterpriseOps-Gym, two benchmarks where the company shows strong results, are aligned with exactly the kind of professional automation scenarios that a recruitment and HR-adjacent platform would generate training signal for. The model card does not specify training data sources, so the extent to which proprietary BOSS Zhipin data influenced post-training remains unstated.

This context matters for enterprise evaluation of the model. An organization deploying Nanbeige4.3-33B for general coding agent tasks is adopting a model whose training likely reflects strong signal in workplace automation workflows but whose broader coding generalization relies on benchmarks that have not been independently validated.

Deployment Reality: Four GPUs Required, No Cloud Hosting Available

Nanbeige4.3-33B is distributed under Apache 2.0 license, which permits commercial use without royalty obligations and imposes no user-count restrictions.

The deployment reality is more constrained. The lab recommends tensor-parallel-size 4, meaning the model requires at minimum four GPU cards to run. At BF16 precision, 33.1 billion parameters require approximately 66 gigabytes of weight storage, which exceeds the memory capacity of any single H100 or A100 in standard configurations. In practice, organizations running the model in production will need 4× A100 80GB instances or equivalent — infrastructure that carries significant cost relative to per-token API access for competing models. Deployment requirements and serving configuration are documented in the model card.

As of October 26, no hosted inference provider has announced support for Nanbeige4.3-33B. Qwen3.8-27B is available through Alibaba Cloud, and both Gemma4-31B and Qwen3.8-27B are accessible through multiple hosted endpoints. This deployment gap is the practical barrier to near-term adoption for teams that do not operate their own GPU infrastructure.

The model supports SGLang and vLLM as production inference frameworks — both widely deployed in high-throughput serving environments — so teams that do run self-hosted inference will find solid toolchain support.

The 3B Distillation Target and What Comes Next

The model card explicitly designates Nanbeige4.3-33B as the "flagship teacher model" for the Nanbeige4.3 series. A companion Nanbeige4.3-3B model is planned as the distillation target: the 33B model generates training signal that the smaller model will learn from. If the 3B distillation reproduces a meaningful fraction of the 33B performance on agentic coding benchmarks, it would produce a model small enough to run on consumer hardware or single enterprise-grade GPUs — addressing the deployment constraint directly.

The critical unknown is whether the claimed benchmark advantages — particularly on Terminal-Bench 2.1 and DeepSWE — will hold up to independent evaluation before or alongside the 3B release. Benchmark reproduction by third parties is the event that will most determine whether Nanbeige LLM Lab's self-reported results represent a genuine technical advance or reflect evaluation choices that favor the lab's own model. Independent evaluation of Terminal-Bench scores, in particular, would settle the most consequential competitive claim: whether a 33B open-weights model can reliably execute real-world terminal tasks at a level currently associated with top closed frontier systems.