Overview
Nanbeige4.3-33B is an open model released on October 25, 2026 by Nanbeige LLM Lab, the research group inside the Chinese recruiting platform BOSS Zhipin, under Apache 2.0. It is not aimed at general chat. The positioning is in the name: agentic coding and office-style cowork.
The architecture is 33B dense with no mixture-of-experts routing, so every token runs through all parameters. The cost shows up in memory: BF16 weights are roughly 66GB, and the model card's SGLang and vLLM recipes use four-way tensor parallelism. What you get in return is simple deployment with no expert routing to tune. Context runs to 524,288 tokens.
Its place in the family is explicit: this is the teacher model. The lab's strategy is to train a strong one first, then distil agentic ability into smaller sizes, starting with Nanbeige4.3-3B. A Base checkpoint ships alongside for custom post-training.
Key Features
- 33B dense, no expert routing: Every token runs through all 33B parameters: 52 layers, hidden size 7,168, grouped-query attention with 56 query heads and 8 KV heads. Simple to deploy, expensive in memory.
- 524,288-token context: A little over 512K tokens of context, RoPE theta of 120 million, vocabulary of 166,912. A whole repository fits in one pass.
- 80.9 on Terminal-Bench 2.1: Vendor-reported 80.9 on Terminal-Bench 2.1. The same table lists 65.5 on DeepSWE, 53.0 on NL2Repo and 70.4 on Toolathlon-Verified.
- Office cowork tasks covered: Automation Bench 54.2, JobBench 44.8, DeepPlanning 63.3. On JobBench the model reports beating the baseline in most of the 35 occupations evaluated.
- Three chat template switches: enable_thinking turns reasoning on, and is on by default. preserve_thinking keeps earlier reasoning across turns, recommended for tool use and code agents. Tool calls are best written as XML, with JSON supported for compatibility.
- Base checkpoint and a distillation plan: Nanbeige4.3-33B-Base ships for custom post-training, with a 3B distillation planned next. Weights are BF16 and community quantisations already exist.
Use Cases
- Self-hosted coding agents that need 30B class capacity with long context
- Terminal work and multi-step code tasks
- Office workflow automation across spreadsheets, documents and systems
- Using the 33B model as a teacher for domain distillation
Pros
- Apache 2.0 with no extra terms for commercial use or further training
- Dense architecture keeps deployment simple with no expert routing to tune
- 512K context avoids splitting long-horizon tasks
- A Base checkpoint makes industry post-training straightforward
- Official four-way parallel recipes for both SGLang and vLLM
Pricing
Apache 2.0 licensed, with weights free to download on Hugging Face for commercial use and further training. Self-hosting needs multiple GPUs: BF16 weights are about 66GB and the official recipes use four-way tensor parallelism, while community quantisations reduce the memory footprint.
Summary
The 30B class is the workhorse size for self-hosted agents. Nanbeige4.3-33B puts dense weights, 512K context and Apache 2.0 in one package, and throws in a Base checkpoint.
Two things to keep in view. Every benchmark is vendor-reported with no independent reproduction yet. And BF16 needs four GPUs, so a single-card deployment means a quantised build, whose behaviour on your tasks you should measure yourself. Its real value may be as a teacher that distils agentic ability down to a size that fits one card.
Version History
- Nanbeige4.3-33B released as an open agentic model (2026-10-25): Nanbeige LLM Lab publishes a 33B dense model on Hugging Face under Apache 2.0, built for agentic coding and office cowork, with 524,288-token context and a Base checkpoint alongside. Self-reported scores include 80.9 on Terminal-Bench 2.1, 65.5 on DeepSWE and 54.2 on Automation Bench. BF16 weights are about 66GB with four-way tensor parallelism in the official recipes. A 3B distillation is planned.