A
aimodels44
Guest
Overview
Nanbeige4.3-33B is Nanbeige’s 33B flagship teacher model for the Nanbeige4.3 series, intended for complex coding, tool-using agents, and professional cowork tasks. The model card reports benchmark results across agentic coding, cowork, and reasoning tasks; these are vendor-reported evaluations, not a guarantee of performance on a particular workload. Maintainer: Nanbeige.
Best use cases
• Complex coding and code-agent workflows.
• Multi-turn tool use and agentic tasks.
• Office, professional cowork, and task automation workflows.
• Reasoning and chat, with thinking enabled or disabled through the chat template.
Limitations
The published benchmark results depend on specific harnesses, settings, and task subsets; for example, Claw-Eval covers 157 single-turn tasks, and YC-bench retains only the latest 10 user turns. The sources do not provide hardware requirements, training details, or independent evaluations. The listed Nanbeige4.2 and Nanbeige4.1 models are different versions and scales, so their descriptions should not be treated as specifications or results for this model.
How it compares
The model card compares this model with Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B at their highest thinking effort using the same inference hyperparameters and harness per benchmark. Nanbeige4.3-33B has higher reported scores on most listed benchmarks, but not all: Qwen3.8-27B scores higher on SWE-Bench Pro and GDPval, while Gemma4-31B scores higher on GPQA Diamond. Related models include Nanbeige4.3-33B-Base, Nanbeige4.2-3B, Nanbeige4.1-3B, Nanbeige4-3B-Thinking-2511, and Nanbeige4.2-3B-Base. The supplied research links concern other Nanbeige versions or unrelated models; they do not establish research findings about this exact model.
Technical specifications
• Parameter scale: 33B, as indicated by the model name; the card describes it as a model in the approximately 30B class.
• Task and library: text generation; Transformers.
• Maximum context length: 524,288 tokens (512K).
• Metadata lists Safetensors and Apache-2.0 licensing.
• The card provides benchmark scores but no hardware requirements, training configuration, or measured latency or cost.
Model inputs and outputs
Inputs
Use chat messages with the tokenizer’s chat template.
enable_thinking controls reasoning for the current response and defaults to enabled; preserve_thinking controls retention of prior assistant reasoning in multi-turn conversations. Passing tools enables the tool-use template; the card recommends tool_call_format="xml" and also supports json for compatibility. Suggested settings are temperature 1.0, top_p 0.95, and 65,536 max new tokens for agentic/tool-use tasks; for reasoning and chat, temperature 0.6, top_p 0.95, and 262,144 max new tokens.Outputs
The model generates text responses and supports reasoning and tool-use chat templates. The supplied Transformers example decodes generated token IDs into text. Tool-call formatting can be XML or JSON; the card recommends XML.
Getting started
Code:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = 'Nanbeige/Nanbeige4.3-33B'
tokenizer = AutoTokenizer.from_pretrained(
model_id,
use_fast=False,
trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype='auto',
device_map='auto',
trust_remote_code=True
)
messages = [{'role': 'user', 'content': 'Which number is bigger, 9.11 or 9.8?'}]
inputs = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
return_tensors='pt'
).to(model.device)
output_ids = model.generate(inputs, max_new_tokens=1024)
response = tokenizer.decode(
output_ids[0][inputs.shape[-1]:],
skip_special_tokens=True
)
print(response)Frequently asked questions
Q: What is Nanbeige4.3-33B designed for?
A: It is presented as a flagship teacher model for complex coding, agentic tool use, and professional cowork tasks.
Q: How long is its context window?
A: The model card states a maximum context length of 524,288 tokens.
Q: Can I turn off reasoning?
A: Yes. Set
enable_thinking to False to use non-thinking mode; it is enabled by default.Q: How should I configure thinking retention?
A: The card recommends
preserve_thinking=False for general chat and question answering, and True for multi-turn tool use, office tasks, and code-agent workflows.Q: Which tool-call formats are supported?
A: The card supports XML and JSON tool-call formats and recommends XML.
Q: Which inference frameworks are shown?
A: The model card provides examples for Transformers, SGLang, and vLLM.
Q: What license is listed?
A: The model metadata lists Apache-2.0.
This is a simplified guide to an AI model called Nanbeige4.3-33B maintained by Nanbeige. If you like these kinds of analysis, join AIModels.fyi or follow us on Twitter.