Nanbeige4.3 opens a 33B model for self-hosted coding agents

What this is about

Nanbeige LLM Lab, the research team at recruitment platform BOSS Zhipin, released Nanbeige4.3-33B on October 24, 2026. The weights are available on Hugging Face, and the model card lists Apache 2.0 as the license. The release date can be traced through the official repository’s metadata. For developers, the main attraction is the ability to run a coding assistant on their own infrastructure and shape its workflows themselves.

This does not guarantee a better programmer. The published performance figures come from the lab itself, not an independent Cyber Ivy comparison. Still, accessible weights, a permissive license and documented tool use provide a concrete reason to test the model against an existing solution. The issue is greater operational choice, not a demonstrated replacement for every commercial frontier model.

What Nanbeige4.3-33B actually does

The language model is intended to support multi-step programming and office tasks. A surrounding agent application supplies files, search results or test output, for example. The model then proposes the next step. That environment determines which files it can change and which commands it can execute. The model weights alone do not constitute a fully configured workplace.

According to the official model card, the model supports up to 524,288 context tokens. That describes the quantity of text units it can process, not guaranteed flawless recall of every detail. The documentation provides startup examples for Transformers, SGLang and vLLM. The team recommends XML for tool calls, while also supporting JSON.

A setting called preserve_thinking can retain reasoning from earlier conversation turns. The lab recommends it for longer tool workflows, rather than every ordinary conversation. The 33B model is also intended to teach smaller models later. The mentioned Nanbeige4.3-3B is an announced next stage, not a finished product established by this release.

Why it matters

Teams running an agent through an external service depend on that service’s availability, interfaces and terms. Open weights provide another deployment option. Whether self-hosting is more economical, however, depends on hardware, utilization and maintenance. The reviewed sources do not establish a universal price advantage.

The lab’s comparison table offers a starting point for internal testing: it reports 80.9 for Nanbeige and 73.0 for Qwen3.8-27B on Terminal-Bench 2.1. On SWE-Bench Pro, however, Qwen leads with 61.7 against Nanbeige’s 60.3. These are results from different tests under their respective documented conditions, not percentages of everyday work reliably completed.

Independent publications explainx.ai and AI/TLDR likewise describe the table as provider-reported. They do not confirm their own reproduction. That makes the practical implication fairly narrow: teams can evaluate another openly available option in their existing workflows, rather than immediately switching providers on the strength of a bar chart.

In plain language

Think of a workshop. An external service is a workshop you hire, which determines its own equipment and processes. Open model weights are more like a machine that can stand in your own workshop. You gain influence over who receives access and how work is performed. In return, you take responsibility for setup, maintenance and quality control.

A strong rating in a tool catalog also does not tell you whether the machine will handle your particular workpiece cleanly. Only a trial with the actual material shows whether it fits. For a coding agent, that material consists of your repositories, tests and failure cases.

A practical example

A fictional development team selects 20 previously resolved bug reports from an internal project. It gives the new model and its existing assistant identical starting code and the same tools. Both work exclusively in separate test copies, without access to production.

The team then counts passing tests, necessary human corrections and effort per accepted solution. It also checks whether an apparently successful proposal removed important checks. The number 20 describes only this invented test design; it is not a result from the release. This turns a general benchmark promise into a verifiable decision for a specific workplace.

Scope and limits

• No independent performance confirmation: The provider selected the comparison models and test environments. The hero graphic shows its comparison, not a Cyber Ivy test. AI/TLDR also notes discrepancies in individual cells between the chart and table; the figures in this article follow the model card’s table.
• Self-hosting requires resources: The metadata lists approximately 33.1 billion BF16 parameters. Long contexts require additional memory. The examples using four-way parallelism are not a promise that the model runs usefully on an ordinary laptop.
• Tool access remains a risk: Open weights prevent neither incorrect changes nor unwanted commands. The model card lists English and Chinese; the reviewed sources provide no robust results for German workflows. Those workflows require their own tests and restricted permissions.

SEO & GEO keywords

Nanbeige4.3-33B, Nanbeige LLM Lab, BOSS Zhipin, open model weights, Apache 2.0, coding agents, Hugging Face, Terminal-Bench, local AI, model comparison