---

license: apache-2.0 language:

• en
• zh

library_name: transformers pipeline_tag: text-generation tags:

• llm
• nanbeige

base_model:

• Nanbeige/Nanbeige4-3B-Base

---

Nanbeige4.1-3B GGUF Models

Model Generation Details

This model was generated using llama.cpp at commit 8872ad212.

---

Quantization Beyond the IMatrix

I've been experimenting with a new quantization approach that selectively elevates the precision of key layers beyond what the default IMatrix configuration provides.

In my testing, standard IMatrix quantization underperforms at lower bit depths, especially with Mixture of Experts (MoE) models. To address this, I'm using the --tensor-type option in llama.cpp to manually "bump" important layers to higher precision. You can see the implementation here:
👉 Layer bumping with llama.cpp

While this does increase model file size, it significantly improves precision for a given quantization level.

I'd love your feedback—have you tried this? How does it perform for you?

---

Click here to get info on choosing the right GGUF model format

---

Introduction

Nanbeige4.1-3B is built upon Nanbeige4-3B-Base and represents an enhanced iteration of our previous reasoning model, Nanbeige4-3B-Thinking-2511, achieved through further post-training optimization with supervised fine-tuning (SFT) and reinforcement learning (RL). As a highly competitive open-source model at a small parameter scale, Nanbeige4.1-3B illustrates that compact models can simultaneously achieve robust reasoning, preference alignment, and effective agentic behaviors.

Specifically, Nanbeige4.1-3B exhibits the following key strengths:

• Strong Reasoning: Nanbeige4.1-3B is capable of solving complex, multi-step problems through sustained and coherent reasoning within a single forward pass, and reliably produces correct final answers on challenging tasks such as LiveCodeBench-Pro, IMO-Answer-Bench, and AIME 2026 I.
• Robust Preference Alignment: Nanbeige4.1-3B achieves solid alignment performance, outperforming not only same-scale models such as Qwen3-4B-2507 and Nanbeige4-3B-2511, but also substantially larger models including Qwen3-30B-A3B and Qwen3-32B on Arena-Hard-v2 and Multi-Challenge.
• Agentic Capability: Nanbeige4.1-3B is the first general small model to natively support deep-search tasks and reliably sustain complex problem solving involving more than 500 rounds of tool invocations. It fills a long-standing gap in the small-model ecosystem where models are typically optimized for either general reasoning or agentic scenarios, but rarely excel at both.

Technical Report: Link

Performances

We evaluate Nanbeige4.1-3B across a broad and diverse set of benchmarks covering general reasoning, and deep-search capabilities.

General Reasoning Tasks

On general reasoning tasks including code, math, science, alignment, and tool-use benchmarks, Nanbeige4.1-3B not only significantly outperforms same-scale models such as Qwen3-4B, but also demonstrates overall superior performance compared to larger models including Qwen3-30B-A3B-2507 and Qwen3-32B.

Benchmark	Qwen3-4B-2507	Qwen3-8B	Qwen3-14B	Qwen3-32B	Qwen3-30B-A3B-2507	Nanbeige4-3B-2511	Nanbeige4.1-3B
Code
Live-Code-Bench-V6	57.4	49.4	55.9	55.7	66.0	46.0	76.9
Live-Code-Bench-Pro-Easy	40.2	41.2	33.0	42.3	60.8	40.2	81.4
Live-Code-Bench-Pro-Mediium	5.3	3.5	1.8	3.5	3.5	5.3	28.1
Math
AIME 2026 I	81.46	70.42	76.46	75.83	87.30	84.1	87.40
HMMT Nov	68.33	48.33	56.67	57.08	71.25	66.67	77.92
IMO-Answer-Bench	48.00	36.56	41.81	43.94	54.34	38.25	53.38
Science
GPQA	65.8	62.0	63.38	68.4	73.4	82.2	83.8
HLE (Text-only)	6.72	5.28	7.00	9.31	11.77	10.98	12.60
Alignment
Arena-Hard-v2	34.9	26.3	36.9	56.0	60.2	60.0	73.2
Multi-Challenge	41.14	36.30	36.97	38.72	49.40	41.20	52.21
Tool Use
BFCL-V4	44.87	42.20	45.14	47.90	48.6	53.8	56.50
Tau2-Bench	45.9	42.06	44.96	45.26	47.70	41.77	48.57

Deep Search Tasks

As a general small model, Nanbeige4.1-3B achieves deep-search performance comparable to specialized agents under 10B parameters.
In contrast to existing small general models, which typically exhibit little to no deep-search capability, Nanbeige4.1-3B represents a substantial qualitative improvement over prior small general models.

#### Deep Search and Agent Benchmarks

Model	xBench-DeepSearch-2505	xBench-DeepSearch-2510	Browse-Comp	Browse-Comp-ZH	GAIA (Text-only)	HLE	SEAL-0
Search-Specialized Small Agents
MiroThinker-v1.0-8B	61	–	31.1	40.2	66.4	21.5	40.4
AgentCPM-Explore-4B	70	–	25.0	29.0	63.9	19.1	40.0
Large Foundation Models (with Tools)
GLM-4.6-357B	70	–	45.1	49.5	71.9	30.4	–
Minimax-M2-230B	72	–	44.0	48.5	75.7	31.8	–
DeepSeek-V3.2-671B	71	–	67.6	65.0	63.5	40.8	38.5
Small Foundation Models (with Tools)
Qwen3-4B-2507	34	5	1.57	7.92	28.33	11.13	15.74
Qwen3-8B	31	2	0.79	5.15	19.53	10.24	6.34
Qwen3-14B	34	9	2.36	7.11	30.23	10.17	12.64
Qwen3-32B	39	8	3.15	7.34	30.17	9.26	8.15
Qwen3-30B-A3B-2507	25	10	1.57	4.12	31.63	14.81	9.24
Ours (with Tools)
Nanbeige4-3B-2511	33	11	0.79	3.09	19.42	13.89	12.61
Nanbeige4.1-3B	75	39	19.12	31.83	69.90	22.29	41.44

Quickstart

For inference hyperparameters, we recommend the following settings:

• Temperature: 0.6
• Top-p: 0.95
• Repeat penalty: 1.0
• Max New Tokens: 131072

For the chat scenario:

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(
  'Nanbeige/Nanbeige4.1-3B',
  use_fast=False,
  trustremotecode=True
)
model = AutoModelForCausalLM.from_pretrained(
  'Nanbeige/Nanbeige4.1-3B',
  torch_dtype='auto',
  device_map='auto',
  trustremotecode=True
)
messages = [
  {'role': 'user', 'content': 'Which number is bigger, 9.11 or 9.8?'}
]
prompt = tokenizer.applychattemplate(
  messages,
  addgenerationprompt=True,
  tokenize=False
)
inputids = tokenizer(prompt, addspecialtokens=False, returntensors='pt').input_ids
outputids = model.generate(inputids.to('cuda'), eostokenid=166101)
resp = tokenizer.decode(outputids[0][len(inputids[0]):], skipspecialtokens=True)
print(resp)

For the tool use scenario:

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(
  'Nanbeige/Nanbeige4.1-3B',
  use_fast=False,
  trustremotecode=True
)
model = AutoModelForCausalLM.from_pretrained(
  'Nanbeige/Nanbeige4.1-3B',
  torch_dtype='auto',
  device_map='auto',
  trustremotecode=True
)
messages = [
    {'role': 'user',  'content': 'Help me check the weather in Beijing now'}
]
tools = [{'type': 'function',
  'function': {'name': 'SearchWeather',
   'description': 'Find out the current weather in a place on a certain day.',
   'parameters': {'type': 'dict',
    'properties': {'location': {'type': 'string',
      'description': 'A city in China.'},
    'required': ['location']}}}}]
prompt = tokenizer.applychattemplate(
  messages,
  tools,
  addgenerationprompt=True,
  tokenize=False
)
inputids = tokenizer(prompt, addspecialtokens=False, returntensors='pt').input_ids
outputids = model.generate(inputids.to('cuda'), eostokenid=166101)
resp = tokenizer.decode(outputids[0][len(inputids[0]):], skipspecialtokens=True)
print(resp)

For the deep-search scenario:

• Inference Framework: miroflow-framework!
• Switch tokenizer configuration to tokenizerconfig_search.json
• Tools Configuration:

Server	Description	Tools Provided
tool-python	Execution environment and file management (E2B sandbox)	createsandbox, runcommand, runpythoncode, uploadfilefromlocaltosandbox, downloadfilefromsandboxtolocal, downloadfilefrominternettosandbox
searchandscrapewebpage	Google search via Serper API	google_search
jinascrapellmsummary	Web scraping with LLM-based information extraction with Jina	scrapeandextract_info

• Summary model: Qwen3-14B-thinking
• Temperature: 1.0
• Note, access to HuggingFace has been explicitly disabled in these tools.

Limitations

While we place great emphasis on the safety of the model during the training process, striving to ensure that its outputs align with ethical and legal requirements, it may not completely avoid generating unexpected outputs due to the model's size and probabilistic nature. These outputs may include harmful content such as bias or discrimination. Please don't propagate such content. We do not assume any responsibility for the consequences resulting from the dissemination of inappropriate information.

Contact

If you have any questions, please raise an issue or contact us at [email protected].

---

🚀 If you find these models useful

Help me test my AI-Powered Quantum Network Monitor Assistant with quantum-ready security checks:

👉 Quantum Network Monitor

The full Open Source Code for the Quantum Network Monitor Service available at my github repos ( repos with NetworkMonitor in the name) : Source Code Quantum Network Monitor. You will also find the code I use to quantize the models if you want to do it yourself GGUFModelBuilder

💬 How to test:
Choose an AI assistant type:
- TurboLLM (GPT-4.1-mini)
- HugLLM (Hugginface Open-source models)
- TestLLM (Experimental CPU-only)

What I’m Testing

I’m pushing the limits of small open-source models for AI network monitoring, specifically:

• Function calling against live network services
• How small can a model go while still handling:

- Automated Nmap security scans - Quantum-readiness checks - Network Monitoring tasks

🟡 TestLLM – Current experimental model (llama.cpp on 2 CPU threads on huggingface docker space):

• ✅ Zero-configuration setup
• ⏳ 30s load time (slow inference but no API costs) . No token limited as the cost is low.
• 🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!

Other Assistants

🟢 TurboLLM – Uses gpt-4.1-mini :

• It performs very well but unfortunatly OpenAI charges per token. For this reason tokens usage is limited.
• Create custom cmd processors to run .net code on Quantum Network Monitor Agents
• Real-time network diagnostics and monitoring
• Security Audits
• Penetration testing (Nmap/Metasploit)

🔵 HugLLM – Latest Open-source models:

• 🌐 Runs on Hugging Face Inference API. Performs pretty well using the lastest models hosted on Novita.

💡 Example commands you could test:

1. "Give me info on my websites SSL certificate"
2. "Check if my server is using quantum safe encyption for communication"
3. "Run a comprehensive security audit on my server"
4. '"Create a cmd processor to .. (what ever you want)" Note you need to install a Quantum Network Monitor Agent to run the .net code on. This is a very flexible and powerful feature. Use with caution!

Final Word

I fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is open source. Feel free to use whatever you find helpful.

If you appreciate the work, please consider buying me a coffee ☕. Your support helps cover service costs and allows me to raise token limits for everyone.

I'm also open to job opportunities or sponsorship.

Thank you! 😊