---
title: "Nanbeige releases a 33B agent model, while smaller successors remain planned"
description: "Nanbeige4.3-33B arrived October 24. What is available, what remains planned, and why download size and vendor benchmarks need careful interpretation."
url: "https://promptszone.com/nanbeige-4-3-33b-open-weights-agent-release"
published: "2026-10-28T06:00:05Z"
categories: ["New Models"]
---

# Nanbeige releases a 33B agent model, while smaller successors remain planned

Nanbeige's October 24 open-weight release offers another coding-agent option. Download size, vendor benchmarks and the smaller-model roadmap need separate scrutiny.

![CSIRO GPU cluster photographed in April 2010; archival computing context, not Nanbeige infrastructure or recommended hardware.](https://promptszone.com/assets/uploads/c65e57dd047b6e3daeec3e16839af630.jpg)

**Summary:**

Nanbeige4.3-33B is available to download; planned smaller successors are not. The release deserves evaluation without treating lab benchmarks as deployment guarantees.

Nanbeige has made its Nanbeige4.3-33B coding and workplace-agent model available as downloadable weights. The [official repository history](https://huggingface.co/Nanbeige/Nanbeige4.3-33B/commits/main) records the initial upload on October 24, 2026. Later documentation and evaluation updates should not be mistaken for a second launch.

The release adds another self-hosting option for developers, but the smaller models described in the lab's roadmap are not part of this delivery. That distinction is central to understanding the announcement: a system intended to help train future compact agents is not itself proof that those agents are ready.

## A teacher today, smaller agents later

The [model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) describes a 33-billion-parameter teacher model, with a planned 3B successor and up to 524,288 tokens of context. Weights are labeled Apache 2.0. Its benchmark table reports gains over similarly sized rivals on several agent tasks, but not every test: Qwen3.8-27B scores higher on SWE-Bench Pro, for example. These are the lab's results, not PromptsZone testing.

A teacher model can supply outputs or other signals used to train a smaller student. For a buyer, the useful question is therefore not just whether the teacher performs well, but whether a future student retains the abilities needed for a particular job. That second claim requires evidence from the student itself.

## Downloadable does not mean effortless to run

The [official file listing](https://huggingface.co/Nanbeige/Nanbeige4.3-33B/tree/main) totals approximately 66.3 GB, with fourteen weight shards alongside tokenizer and configuration files. That download size is not a verified runtime-memory requirement. Serving also needs working memory, and the chosen context length and concurrency can change the resource budget substantially.

Teams considering deployment should separate three decisions: whether they may use the weights under the stated terms, whether their infrastructure can serve the workload, and whether the model is reliable enough with their tools. A benchmark score answers none of those questions by itself. A sensible evaluation would include failed tool calls, interrupted workflows and recovery, not only successful demonstrations.

## Why the release is worth watching

Nanbeige comes from the company behind recruitment platform BOSS Zhipin. [Kanzhun's March 2026 results announcement](https://ir.zhipin.com/news-releases/news-release-details/kanzhun-limited-announces-fourth-quarter-and-full-year-2025/) identifies Nanbeige as its self-developed model and discusses AI use in recommendations and recruiting products. That background helps explain the interest in workplace tasks; it does not establish this release's performance in hiring or justify automated employment decisions.

The near-term opportunity is a new model to evaluate under one's own operating conditions. The longer-term promise remains smaller agents trained from it. Keeping those two stages separate avoids turning a development plan into an availability claim.

Photo: [CSIRO GPU cluster, April 19, 2010](https://commons.wikimedia.org/wiki/File:CSIRO_ScienceImage_11313_The_CSIRO_GPU_cluster_at_the_data_centre.jpg), CSIRO, [CC BY 3.0](https://creativecommons.org/licenses/by/3.0/), unmodified. Archival computing context only, not Nanbeige infrastructure or a recommended deployment configuration.