Hai-2.6-Extend-0913-GGUF

H.ai 2.6 Extend 0913

<p align="center"> <img src="https://img.shields.io/badge/Base-Qwen3.5--9B-blue" alt="Base Model"> <img src="https://img.shields.io/badge/Format-GGUF-green" alt="Format"> <img src="https://img.shields.io/badge/Mirror-ModelScope-purple" alt="Mirror"> </p>

H.ai — a gentle, patient AI companion that walks you through things one step at a time.

This repository is the HuggingFace mirror of Hai-2.6-Flash-Extend. Its contents are identical to the official ModelScope repository. The Reinforced build improves character consistency and mathematical ability over Hai-2.6-Flash.

---

📋 Model Family

The H.ai series is trained by Horn.Studio. The primary release platform is ModelScope; below is the index of all previous versions:

The 2.6 series is the successor product line to Hai-2-Series. Hai-3 will use a new base model.

---

📋 Model Information

---

✨ Highlights

Persona

• Gentle and patient: never impatient, always willing to break complex concepts down into plain language
• Teaching-friendly: a typical non-think CoT model (the think tags are retained)
• High-school subject tutoring: deeply aligned for Chinese composition, mathematics, English continuation writing and composition, physics, chemistry, and politics — accuracy preserved while making explanations more detailed and approachable
• Character consistency: the Reinforced build specifically strengthens persona consistency across long conversations
• Code-friendly: inherits the strong coding ability of the Qwen base model

Capabilities

• ✅ Role-play and emotional companionship
• ✅ Everyday conversation and emotional support
• ✅ Creative writing and copy polishing
• ✅ Programming tutoring (Python / Java / C++ / front-end, etc.)
• ✅ Technical concept breakdown (AI, algorithms, system architecture, etc.)
• ✅ Mathematical reasoning and logical analysis
• ✅ Agent ability restored to over 80% of Qwen3.5 9B

Per-Batch Behavior Metrics

---

🛡️ Safety Guardrails

We used the heretic tool to strip the safety layer from a Hai2 model, producing Hai-2-Abliterated, and then fully restored all of the original model's safety guardrails through full fine-tuning — making them more robust than before. Experiments show that an Abliterated model can, under certain conditions, re-learn safety guardrails and achieve more solid results. This model uses that method to strengthen its guardrails.

---

🚀 Quick Start

Option 1: llama.cpp (recommended, local)

Download the quantization you want and load it directly:

# Q4_K_M (~5.5 GB, recommended)
./llama-cli -m Hai-2.6-Flash-Extend.Q4_K_M.gguf -p "Hello, H.ai"

# Q6_K (~7.2 GB, higher quality)
./llama-cli -m Hai-2.6-Flash-Extend.Q6_K.gguf -p "Hello, H.ai"

Option 2: LM Studio / Ollama

1. Download the .gguf file locally
2. LM Studio: drop it into the models directory and it will be detected automatically. Leaving the System Prompt empty is recommended, and a relatively high temperature works best
3. Ollama: create a Modelfile

FROM ./Hai-2.6-Flash-Extend.Q4_K_M.gguf

PARAMETER temperature 0.7
PARAMETER top_p 0.9

ollama create hai -f Modelfile
ollama run hai

Option 3: llama-cpp-python (Python)

from llama_cpp import Llama

llm = Llama(
    model_path="./Hai-2.6-Flash-Extend.Q4_K_M.gguf",
    n_ctx=8192,
    chat_format="chatml",
)

resp = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Hello, H.ai"}],
    temperature=0.7,
    top_p=0.9,
)
print(resp["choices"][0]["message"]["content"])

---

📦 Files

This repository provides the following formats:

Note: The --allow-requantize variants are re-quantized after merging, so some layers go through multiple quantization rounds. In extreme cases they differ slightly from a native Q4_K_M, but for everyday conversation the difference is imperceptible.

---

⚙️ Training Configuration (Reference)

The following is the training configuration of H.ai 2.0 (27B), provided as a reference for the technical approach behind the whole series:

{
  "base_model": "Qwen/Qwen3.5-27B",
  "quantization": "4-bit (bnb NF4)",
  "lora_r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
  "learning_rate": 1e-4,
  "batch_size": 1,
  "gradient_accumulation_steps": 8,
  "epochs": 2,
  "warmup_steps": 20,
  "lr_scheduler": "cosine",
  "weight_decay": 0.01,
  "max_seq_length": 4096,
  "packing": true
}

---

⚠️ Known Issues

• The persona of this build appears to have a problem: it tends to talk back. We are investigating the cause.

---

⚠️ Usage Notes

1. Base license: This model is fine-tuned from Qwen3.5-9B and follows its original license.
2. Commercial use: Please consult the latest Qwen3.5 license terms to confirm commercial compliance.
3. Content safety: We do not guarantee the safety of model outputs, as this safety layer has not yet undergone complete validation and it is uncertain whether its effect generalizes beyond the dataset.
4. Disclaimer: This model is provided "as is"; the authors accept no responsibility for the results of its use.

---

🙏 Acknowledgements

• Qwen for the powerful base model
• Unsloth for the efficient fine-tuning framework
• llama.cpp for GGUF conversion and inference tooling
• heretic for the safety-layer ablation tool

---

📬 Feedback

ModelScope is the primary release platform. For questions or suggestions, please reach out on the ModelScope discussion board first:

• ModelScope: Horn.Studio profile ｜ original repository of this model
• HuggingFace discussions / issues are also welcome (mirror sync may lag)

Enjoy chatting with H.ai~ (´∇ﾉ｀*)ノ