Q

Qwen3.5-27B-BlueStar-Derestricted-Lite

RP Chat

No ratings yet(0)
Prose
—(0)
Intelligence
—(0)
Knowledge
—(0)
Instruction Following
—(0)
Coding
—(0)
Steerability
—(0)

Ratings reflect this model as run on Arli AI.

Qwen3.5-27B-BlueStar-Derestricted-Lite is a text generation model available on Arli AI. It offers 27B parameters, 262K token context, reasoning support, vision input. Recommended use: RP Chat. Run it through the OpenAI-compatible chat completions API with unlimited requests and zero-log privacy.

View on Hugging FaceBack to Models

Hourly Usage

Token Usage

Input (24H)

273

Output (24H)

97

Input (7D)

17K

Output (7D)

105

Performance Metrics

Typical Total Time

3.01s

Typical TTFT

0.95s

Typical Input TPS

1296.97

Typical Output TPS

26.99

Model Information

Context Size

262144

Quantization

r64 on INT8

Engine

vllm

Creation Method

Finetune

Model Class

Qwen3.5-27B-D - LoRA-DRS-Lite

Chat Template

Qwen3.5

Reasoning

Yes

Vision

Yes

Parameters

27B

Added At

4/4/2026

Use Qwen3.5-27B-BlueStar-Derestricted-Lite with the Arli AI API

Qwen3.5-27B-BlueStar-Derestricted-Lite is available through an OpenAI-compatible chat completions endpoint. Point any OpenAI SDK at https://api.arliai.com/v1, pass Qwen3.5-27B-BlueStar-Derestricted-Lite as the model, and requests are served without logging.

cURL

curl https://api.arliai.com/v1/chat/completions \
  -H "Authorization: Bearer $ARLIAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3.5-27B-BlueStar-Derestricted-Lite",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.arliai.com/v1",
    api_key="YOUR_ARLIAI_API_KEY",
)

response = client.chat.completions.create(
    model="Qwen3.5-27B-BlueStar-Derestricted-Lite",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

See the text generation docs for every parameter, or read the quick start guide to make your first request.


license: mit datasets:

  • zerofata/Instruct-Anime
  • zerofata/Gemini-3.1-Pro-SmallWiki
  • zerofata/Gemini-3.1-Pro-GLM5-Characters
  • zerofata/Roleplay-Anime-Characters base_model:
  • Qwen/Qwen3.5-27B

BlueStar
image

BlueStar v1

Qwen3.5 27B
01 Overview

An experimental tune on Qwen 3.5 27B.

Designed for conversational assistant tasks and RP.

Model feels pretty creative and has some nice moments. Couple brainfarts and bits of repetition occasionally, but nothing out of the normal. (The qwen team themselves recommend a presence penalty of 1.5. Yikes)

Non thinking and thinking are both supported. Thinking has reduced censorship as the original thinking refusals didn't seem to generalize well to the new format I gave it.

02 SillyTavern Settings
Recommended Roleplay Format
ActionsIn plaintext
Dialogue"In quotes"
Thoughts*In asterisks*
Recommended Samplers
Temp0.8
MinP0.05 - 0.075
Rep Pen1.00 - 1.1
03 Quantizations
GGUF
iMatrix
04 Creation Process

Creation Process: SFT

SFT on approx 23 million tokens (12 million trainable). New is some Gemini Synth data which replaces some of my lower quality datasets.

About 10% of the dataset included reasoning for creative assistant tasks. This reasoning seems to have generalized quite well to other parts of the model and heavily reduces the token usage of thinking.

I think this model still needs a pass over with DPO to try and tackle the repetition and some of the weird oddities of the original instruct model, but that'll need to wait. I've been overspending training these models recently.

Trained using MS-Swift.

MS-Swift Config
SFT (4*H200)
PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
USE_HF=True \
WANDB_PROJECT=Qwen3.5-27B-SFT \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
NPROC_PER_NODE=4 \
swift sft \
    --model Qwen/Qwen3.5-27B \
    --tuner_type lora \
    --dataset '/workspace/think_dataset.jsonl' \
              '/workspace/nothink_dataset.jsonl' \
    --torch_dtype bfloat16 \
    --bf16 true \
    --use_liger_kernel true \
    --lora_rank 128 \
    --lora_alpha 16 \
    --use_rslora true \
    --target_modules all-linear \
    --freeze_llm false \
    --freeze_vit true \
    --freeze_aligner true \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 16 \
    --num_train_epochs 2 \
    --learning_rate 2e-5 \
    --warmup_ratio 0.05 \
    --max_length 10752 \
    --split_dataset_ratio 0.01 \
    --add_non_thinking_prefix true \
    --load_from_cache_file true \
    --group_by_length true \
    --eval_steps 200 \
    --save_steps 200 \
    --save_total_limit 10 \
    --logging_steps 1 \
    --dataloader_num_workers 8 \
    --output_dir output/Qwen3.5-27B-SFT-Model \
    --report_to wandb