What is Nanbeige4.3-33B?
Nanbeige4.3-33B is a dense 33-billion-parameter text model from the Nanbeige LLM Lab at BOSS Zhipin, the Chinese recruitment platform run by Kanzhun. It was released on Hugging Face on 24 October 2026 with a separate base model for custom post-training. It reads English and Chinese, has a 524,288-token (512K) context and a thinking mode that is on by default.
The lab describes it as a teacher model: it was built to excel at coding agents and office-style "cowork" tasks, and will be distilled into a forthcoming Nanbeige4.3-3B. The chat template supports XML or JSON tool calls and an option to keep earlier reasoning across turns.
The weights are under Apache 2.0. There is no hosted API from the lab; the card gives SGLang and vLLM recipes using four-way tensor parallelism.
How much does Nanbeige4.3-33B cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0 |
| Hosted API | Not offered | No first-party endpoint listed |
Open weights under Apache 2.0; no official hosted API.
Is Nanbeige4.3-33B worth it?
Nanbeige4.3-33B's claims are large and entirely self-reported. In the lab's table it beats Qwen3.8-27B on 18 of 20 benchmarks, with the widest gaps on agentic coding: 65.5 against 42.5 on DeepSWE, 23.8 against 6.3 on Terminal-Bench 4.0 and 80.9 against 73.0 on Terminal-Bench 2.1. It trails Qwen on SWE-Bench Pro (60.3 to 61.7) and GDPval (84.8 to 86.4). The lab says every model ran with the same harness and settings, but a week after release no independent leaderboard lists it.
Some of the harness choices deserve a second look. Its own scores were taken with preserved thinking switched on, SWE-Atlas answers were graded by GPT-6 Sol, and several cowork benchmarks are young enough that small changes in setup move results a lot. The pattern of very large leads on new agent benchmarks and parity on established ones is common in launch tables.
The practical trade-off is architecture. Unlike the mixture-of-experts models it competes with, all 33 billion parameters run on every token. BF16 weights need about 66GB, and one analysis puts a 4-bit build at about 21GB plus 28GB of cache at 128K context, beyond a single consumer GPU. It is also text-only, where Qwen3.8-27B reads images.
What it does offer is a clean Apache 2.0 licence, a released base model and a 512K context, which together make it a reasonable starting point for a company wanting to post-train its own agent. As an off-the-shelf choice it needs independent numbers before it displaces Qwen.
Where it earns its place
A platform team at a 200-engineer software company
Run an internal coding agent on its own GPUs
Serve it with the card's vLLM recipe on four GPUs with tensor parallelism and the qwen3_coder tool parser. Benchmark it against Qwen3.8-27B on the team's own repositories first, since the 23-point DeepSWE lead is the lab's figure.
An ML lead at a vertical-SaaS firm
Post-train a domain agent for HR or finance workflows
Start from Nanbeige4.3-33B-Base under Apache 2.0 with no revenue caps. The lab invites joint work on industry agents, which may help with data but means no published support terms.
A developer evaluating open models for long-context office tasks
Process long documents and multi-step tool use in one context
The 512K context fits whole contracts or codebases, but at 128K the cache alone is about 28GB per one estimate, so the recommended 65,536-token output budget for agent tasks has a real memory cost.
Before you commit
- Setup
- Download the weights and serve with SGLang or vLLM using --tp-size 4 and the qwen3 reasoning and qwen3_coder tool parsers; Transformers needs trust_remote_code.
- Integrations
- Transformers, SGLang, vLLM; OpenAI-compatible endpoints from either server; the lab ran NL2Repo and JobBench through the Claude Code harness.
- Data and privacy
- Apache 2.0 with no use restrictions; self-hosted only, so no third-party data terms apply. No data-provenance statement beyond the model card.
- Cost at scale
- BF16 weights need about 66GB, so four 24GB or two 48GB GPUs before cache; a 4-bit build is about 21GB plus about 28GB of cache at 128K context.
- Limits
- Text only; English and Chinese; recommended output budgets of 65,536 tokens for agents and 262,144 for reasoning make long runs slow on a dense model.
What’s good
- Leads Qwen3.8-27B on 18 of 20 of the lab's benchmarks, widest on agentic coding
- Apache 2.0, with the base model released for post-training
- 512K-token context
- XML and JSON tool-call formats with optional preserved reasoning
What’s not
- All benchmark results are self-reported; no independent ranking yet
- Dense 33B: about 66GB in BF16, slower per token than 3B-active MoE rivals
- Text only; Qwen3.8-27B also reads images
- No hosted API from the lab
Best for
Teams that self-host and want an Apache-licensed open coding agent, or a base model to post-train, and can verify its claims on their own tasks.
Nanbeige4.3-33B alternatives
Nanbeige4.3-33B: common questions
- Is Nanbeige4.3-33B worth it?
- On its makers' numbers Nanbeige4.3-33B is the strongest open coding agent near 30B, but no one outside the lab has confirmed that yet, and being dense makes it heavy to serve. We rate it 3.8 out of 5.
- How much does Nanbeige4.3-33B cost?
- Open weights under Apache 2.0; no official hosted API.
- What are the downsides of Nanbeige4.3-33B?
- All benchmark results are self-reported; no independent ranking yet; dense 33B: about 66GB in BF16, slower per token than 3B-active MoE rivals; text only; Qwen3.8-27B also reads images; no hosted API from the lab.
- Who is Nanbeige4.3-33B best for?
- Teams that self-host and want an Apache-licensed open coding agent, or a base model to post-train, and can verify its claims on their own tasks.
- How much does Nanbeige4.3-33B cost for a team or in production?
- BF16 weights need about 66GB, so four 24GB or two 48GB GPUs before cache; a 4-bit build is about 21GB plus about 28GB of cache at 128K context.
- Is Nanbeige4.3-33B safe for company data?
- Apache 2.0 with no use restrictions; self-hosted only, so no third-party data terms apply. No data-provenance statement beyond the model card.
- What are the best Nanbeige4.3-33B alternatives?
- The closest alternatives we have reviewed are Qwen3.8-27B, Gemma 4, Granite 4.2 and Sage-A1.5.
Answers checked Oct 31, 2026.