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?
Long-context reasoning is a critical capability for modern Large Language Models (LLMs), yet a significant gap exists in the post-training stage. The field lacks a mature, end-to-end system that provides: (1) a scalable pipeline for synthesizing challenging long-context reasoning data, (2) reinforcement learning (RL) methods tailored to the nuances of long-context reasoning, and (3) agent architectures designed to operate on information streams that exceed the model's context capacity.
In this work, we introduce QwenLong-L1.5, a long-context reasoning model built upon Qwen3-30B-A3B-Thinking, augmented with memory mechanisms to process tasks far beyond its physical context window. Our core contribution is a full post-training recipe that unifies data synthesis, training methodologies, and agent architectures.
First, to address the scarcity of high-quality data, we developed a novel synthesis pipeline that moves beyond simple "needle-in-a-haystack" tasks. Instead, it focuses on creating challenges that require multi-hop grounding and reasoning over globally distributed evidence. This is achieved by deconstructing source documents into atomic facts and programmatically composing complex, verifiable questions from this structured information.
Second, we introduce several RL strategies to tackle the instability of long-context training. This includes task-balanced sampling to stabilize mini-batch distributions and our novel Adaptive Entropy-Controlled Policy Optimization (AEPO) algorithm, which employs an entropy-based mechanism to actively control gradients and sustain training on sequences of progressively increasing length.
Third, to handle tasks exceeding the model's physical window, we introduce a memory management framework. Through a multi-stage fusion RL paradigm, we synergistically combine the model's single-pass reasoning (within its 256K window) with an iterative memory updating mechanism to extend its operational range.
Our comprehensive evaluation on leading long-context benchmarks shows that QwenLong-L1.5 surpasses its Qwen3-30B-A3B-Thinking baseline by an average of 9.9 points, achieving performance comparable to top-tier models like GPT-5 and Gemini-2.5-Pro. Crucially, these enhancements also translate to significant gains in general domains like mathematics, tool-use, and long-dialogue scenarios, demonstrating that strong long-context ability provides a foundational boost to a model's overall reasoning capabilities.
🛠️ Requirements
bash
1# Create the conda environment2conda create -n qwenlongl1_5 python==3.103conda activate qwenlongl1_5
45# Install requirements6pip3 install -r requirements.txt
78# Install verl, we use the 0.4 version of verl9git clone --branch v0.4 https://github.com/volcengine/verl.git
10cd verl
11pip3 install -e .
🚀 Quick Start
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
23model_name ="Tongyi-Zhiwen/QwenLong-L1.5-30B-A3B"45# load the tokenizer and the model6tokenizer = AutoTokenizer.from_pretrained(model_name)7model = AutoModelForCausalLM.from_pretrained(8 model_name,9 torch_dtype="auto",10 device_map="auto"11)1213# prepare the model input14template ="""Please read the following text and answer the question below.
1516<text>
17$DOC$
18</text>
1920$Q$
2122Format your response as follows: "Therefore, the answer is (insert answer here)"."""23context ="<YOUR_CONTEXT_HERE>"24question ="<YOUR_QUESTION_HERE>"25prompt = template.replace('$DOC$', context.strip()).replace('$Q$', question.strip())26messages =[27# {"role": "system", "content": "You are QwenLong-L1, created by Alibaba Tongyi Lab. You are a helpful assistant."}, # Use system prompt to define identity when needed.28{"role":"user","content": prompt}29]30text = tokenizer.apply_chat_template(31 messages,32 tokenize=False,33 add_generation_prompt=True34)35model_inputs = tokenizer([text], return_tensors="pt").to(model.device)3637# conduct text completion38generated_ids = model.generate(39**model_inputs,40 max_new_tokens=50000,41 temperature=0.7,42 top_p=0.9543)44output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()4546# parsing thinking content47try:48# rindex finding 151649 (</think>)49 index =len(output_ids)- output_ids[::-1].index(151668)50except ValueError:51 index =05253thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")54content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")5556print("thinking content:", thinking_content)57print("content:", content)
📝 Citation
If you find this work is relevant with your research or applications, please feel free to cite our work!
@article{shen2025qwenlongl15,
title={QwenLong-L1.5: Post-Training Recipe for Long-Context Reasoning and Memory Management},
author={Weizhou Shen and Ziyi Yang and Chenliang Li and Zhiyuan Lu and Miao Peng and Huashan Sun and Yingcheng Shi and Shengyi Liao and Shaopeng Lai and Bo Zhang and Dayiheng Liu and Fei Huang and Jingren Zhou and Ming Yan},
journal={arXiv preprint arXiv:2512.12967},
year={2025}
}
🚀 If you find these models useful
Help me test my AI-Powered Quantum Network Monitor Assistant with quantum-ready security checks:
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:
"Give me info on my websites SSL certificate"
"Check if my server is using quantum safe encyption for communication"
"Run a comprehensive security audit on my server"
'"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.
