Qwen3.8-27B-1C

Qwen3.8-27B-1C

Helps developers on the 1C:Enterprise platform by reading and fixing BSL code and finding objects in XML exports. A Russian-first fine-tune of Qwen3.8-27B, offered here as GGUF.

Made by sweetand on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3.8-27B-1C is a supervised fine-tune of Qwen3.8-27B for development on the 1C:Enterprise platform. It focuses on BSL code, the BSL plus XML layout of configuration exports, XML schemas and common configuration practice. Russian is the main language, with English also supported. It is a LoRA adapter, and this repository lists GGUF files.

Who it suits: 1C developers who want an assistant that knows the platform's code style and its project layout.

Industry | Who uses it | How they use it | Example
Software | 1C developers | Ask where objects and modules sit in an XML export | Find where a catalog's object module lives
Software | Code reviewers | Review and explain BSL written by others | Get a walkthrough of an unfamiliar module
Enterprise IT | Platform teams | Refactor BSL to follow house style | Rename procedures to match their role

If you do not work with 1C, you can skip it, since its strength is limited to that domain.

Where to use it:

• Your own computer with Python and the transformers library
• llama.cpp or another GGUF runtime, after you export or merge the adapter
• An OpenAI-compatible endpoint that you serve yourself

Use case | Who | What they get back
Project navigation | 1C developers | Where objects and modules sit in an XML export
BSL writing and refactoring | Developers | Code in 1C style
Code review | Reviewers | Explanations of platform mechanisms and others' BSL

How to set it up:

• Python with pip
• A GPU, with 13.7 GB of space for the 4-bit files as a rough figure
• The base model's chat template, which the card says to use

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load them as shown below, replacing the placeholder repository name with the real one.

pip install -U huggingface_hub
hf download sweetand/qwen3.8-27b-1C

How to use it: Apply the Qwen3.8 chat template and start from the sampling values the card gives for precise coding. Thinking modes and tool calling come from the base Qwen3.8 model.

Setting | Value
temperature | 0.6
top_p | 0.95
top_k | 20
min_p | 0.0
presence_penalty | 0.0
repetition_penalty | 1.3
effort | medium

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "YOUR_HF_USER/Qwen3.8-27B-1C"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype="auto", device_map="auto", trust_remote_code=True
)
messages = [
    {"role": "user", "content": "Где в выгрузке конфигурации обычно лежит модуль объекта справочника?"}
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, temperature=0.6, top_p=0.95)
print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Example: You ask, in Russian, where the object module of a catalog usually sits in a configuration export. The model answers in text with the typical location. Check the answer in the configurator, since the card warns about invented paths.

For newcomers and experts: If you are new to 1C, treat this as a helper for reading and drafting code, not as a source of truth. Always test its output in your configurator or CI.

For experts, the training data covered BSL plus XML project structure, platform development standards, XML schema quirks, open 1C code bases, meaningful naming and the 1C query language. The context window is inherited from Qwen3.8, up to 262K tokens, and long contexts depend heavily on hardware.

Qwen3.8-27B is the base model
Supervised fine-tuning on 1C data adds a LoRA adapter
You apply the base chat template
The model answers in 1C style

Good to know: The card names the limits openly.

• License: Apache 2.0, as the Hugging Face record lists it. The card asks you to respect the licenses of open-source code bases used in training.
• Cost to run: no API key or account is named. You need a GPU for a 27B model.
• Data it reaches: downloading reaches Hugging Face. The card does not describe the model reaching files or the web.
• Not for: replacing official 1C documentation or a security audit, a source of truth for closed configurations, or a guarantee that code compiles. It may invent object names, XML paths and typical APIs.

On Hugging Face, 13 November 2026

• Task: image text to text
• Parameters: 27.3B
• Weights: about 13.7 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 36,131
• Likes: 30
• License: apache-2.0
• Languages: ru, en
• Based on: Qwen/Qwen3.8-27B
• Tags: gguf, 1c, onescript, image-text-to-text, conversational

Links

• Model: https://huggingface.co/sweetand/qwen3.8-27b-1C
• Files: https://huggingface.co/sweetand/qwen3.8-27b-1C/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sweetand/qwen3.8-27b-1C

Then follow the usage in its model card. Nothing in this kit runs by itself.