Production-ready · 27B parameters · MIT

QwenFable

A production-grade, fully fine-tuned 27B language model — engineered for advanced reasoning, software engineering, and structured problem solving.

27B
Parameters
105T
Training Tokens
262k
Context Length
Q4_K_M
GGUF Quant
16.5GB
Model Size

About QwenFable

QwenFable (also known as Qwable 27B) is a full fine-tune of unsloth/Qwen3.6-27B on a proprietary 105-trillion-token synthetic corpus generated using Claude Mythos and Fable 5. The result is a model optimized for real-world capability over benchmark scores.

Unlike adapter-based releases, this is a complete merged Hugging Face checkpoint — deployable natively across Transformers, vLLM, TGI, llama.cpp, Ollama, LM Studio, and more.

Capabilities What QwenFable excels at

🧠

Multi-step Reasoning

Decomposes complex problems into clear logical chains before answering.

💻

Software Engineering

Produces production-quality code with strong debugging and architecture skills.

📐

Algorithmic Thinking

Handles algorithm design, system architecture, and technical documentation.

📝

Structured Writing

Generates clear, technically accurate, logically organized responses.

🔧

Instruction Following

Strong adherence to instructions with deterministic output formatting.

🤖

Agent-oriented

Designed for agent workflows, tool calling, and structured interactions.

Quick Start Run QwenFable in one command

bash
# Pull and run with Ollama
ollama run hf.co/JinglanWeb3/QwenFable:Q4_K_M
bash
# Start an OpenAI-compatible server
llama-server -hf JinglanWeb3/QwenFable:Q4_K_M

# Or chat directly in the terminal
llama-cli -hf JinglanWeb3/QwenFable:Q4_K_M
text
1. Open LM Studio
2. Search for "JinglanWeb3/QwenFable"
3. Select the Q4_K_M quant
4. Load and chat
python
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained(
    "JinglanWeb3/QwenFable",
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("JinglanWeb3/QwenFable")

⚠️ You need to accept the gated access terms on the model page first.

Example Try QwenFable via llama-cpp-python

python
import llama_cpp

llm = llama_cpp.Llama.from_pretrained(
    repo_id="JinglanWeb3/QwenFable",
    filename="Qwable-27b_Q4_K_M.gguf",
    n_ctx=8192,
    n_gpu_layers=-1,  # offload all layers to GPU
)

response = llm.create_chat_completion(
    messages=[{
        "role": "user",
        "content": "Write a Python function to merge two sorted lists."
    }]
)
print(response["choices"][0]["message"]["content"])
python
import llama_cpp

llm = llama_cpp.Llama.from_pretrained(
    repo_id="JinglanWeb3/QwenFable",
    filename="Qwable-27b_Q4_K_M.gguf",
    n_ctx=8192,
    n_gpu_layers=-1,
    # Enable reasoning extraction
    chat_format="qwen",
)

# The model outputs reasoning inside  tags
stream = llm.create_chat_completion(
    messages=[{
        "role": "user",
        "content": "Explain quantum entanglement simply."
    }],
    stream=True,
)

for chunk in stream:
    delta = chunk["choices"][0]["delta"]
    if "content" in delta:
        print(delta["content"], end="")

Supported Tools Run QwenFable with your favorite app

Model Specifications

Base Modelunsloth/Qwen3.6-27B
ArchitectureQwen 3.6
Parameters27 Billion
Training MethodFull Supervised Fine-Tuning
Training CorpusClaude Mythos + Fable 5 Synthetic (105T tokens)
Context Length262,144 tokens
PrecisionBF16
GGUF QuantQ4_K_M (16.5 GB)
LicenseMIT
Languagesen, zh, ko, hi, sa, ta, te, fr, es, mr, gd, br

🌟 Want this as a live chat demo?

This Space is currently a static page. With a community GPU grant from Hugging Face, it can become a full interactive ZeroGPU demo. Open a discussion below to apply.

Apply for GPU Grant